
### Qwen Image 2.1: Unified Image Generation and Editing

Qwen Image 2.1 is a unified text-to-image generation and image editing model in the Qwen series. Where earlier releases split generation and editing across Qwen Image and Qwen Image Edit, version 2.1 handles both in a single model. Its visual generation component has just 7B parameters across 32 single-stream DiT layers, so it produces high-quality images at a fraction of the compute of larger diffusion models.

#### Compact and Efficient Architecture

The model uses mixed-granularity attention and prefix KV cache reuse to keep inference cheap without sacrificing output quality. It generates at a native 2048 x 2048 resolution, with supported aspect ratios from 1:1 through 16:9 and 9:16, including 4:3, 3:4, 3:2, and 2:3 formats.

#### Native Transparency

Qwen Image 2.1 generates transparent RGBA images directly from text prompts, so there is no separate background-removal pass. The same model can edit transparent layers and extract subjects from ordinary photographs, turning a product shot or portrait into a clean cutout ready for compositing. Stickers, icons, logos, and design assets come out with a real alpha channel.

#### Versatile Editing

Editing accepts up to 10 reference images in a single request. You can mark local edits by drawing circles, painting annotations, or providing a separate mask, and the model preserves identity for both people and products across edits. This makes it possible to combine several individually photographed portraits into one coherent group shot, or to place a product into a new scene without it drifting from the original.

#### Realistic Textures and Refined Aesthetics

The release improves typography, portrait lighting, and fine surface detail. Text rendering carries over the Qwen series' strength with legible, well-composed lettering, which is useful for signage, posters, and product packaging.

#### Use Cases

- Transparent assets for design systems, stickers, icons, and UI graphics
- Subject extraction and background removal from photographs
- Product photography variants with consistent product identity
- Multi-reference composition, such as group photos assembled from individual portraits
- Localized edits guided by masks or painted annotations
- Posters, signage, and marketing graphics with accurate in-image text

#### Included ComfyUI Workflows

This template ships with three ready-to-run ComfyUI workflows: text-to-image generation, reference-based image editing, and background removal. They use the Comfy-Org INT8 repackage of the diffusion model and the Qwen3-VL 8B text encoder, which keeps memory use low while preserving quality. For the earlier generation of the family, see the Qwen Image and Qwen Image Edit entries.


Compact 7B unified text-to-image and image editing model in the Qwen series with native transparent (RGBA) generation and multi-reference editing

Qwen Image 2.1


### Qwen3.8 27B: Dense Vision-Language Model with Flexible Thinking Control

Qwen3.8 27B is a dense multimodal foundation model from Alibaba's Qwen team, described by its authors as part of "the most capable generation in the Qwen open-model family to date." Built on the Qwen3.5 hybrid Gated DeltaNet and Gated Attention architecture, it is a native vision-language model that understands images and videos while offering per-request control over how much it reasons before answering. At 27 billion parameters it targets coding, professional work, research, and long-horizon agentic tasks.

#### Key Features

-   **Broad Capability Gains** - Comprehensive improvements across coding, professional work, research, and long-horizon agentic tasks
-   **Agent Execution** - Stronger autonomous planning and better handling of environment feedback, designed to carry complex multi-step tasks through to completion
-   **Flexible Thinking Control** - Thinking is on by default and can be disabled per request; reasoning depth is tunable across xhigh, medium, and low, and prior reasoning can be carried forward between turns
-   **Vision-Language Understanding** - Spans STEM diagrams and documents through to hour-scale video comprehension
-   **Downstream Compatibility** - Broader support for popular agent harnesses and development tools
-   **Long Context** - 262,144 tokens natively, extensible up to 1,000,000 tokens with RoPE scaling such as YaRN

#### Architecture

-   Causal Language Model with Vision Encoder
-   27B dense parameters
-   64 layers with a 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)) hybrid layout
-   Gated DeltaNet linear attention (48 V heads, 16 QK heads, head dim 128)
-   Gated Attention (24 Q heads, 4 KV heads, head dim 256, rotary dim 64)
-   Hidden dimension 5120, Feed Forward Network intermediate dimension 17408
-   Multi-token prediction (MTP) trained with multiple steps
-   Native 262K context, extensible to 1M tokens

#### Use Cases

-   Coding and software engineering agents
-   Professional and office workflows carried across many steps
-   Computer, browser, and mobile control agents
-   Document, chart, and STEM diagram analysis
-   Visual web development and application recreation
-   Long-video comprehension and long-context research over large document sets

#### Benchmarks

On the Qwen3.8 benchmark suite ([source](https://huggingface.co/Qwen/Qwen3.8-27B)), Qwen3.8 27B scores SWE-bench Pro 61.7, QwenSWEBench 79.0, LiveCodeBench v6 90.3, GPQA Diamond 89.2, IFBench 79.5, CoWorkBench 70.7, and HLE 30.8 on text tasks. On multimodal tasks it scores OSWorld-Verified 84.3, AndroidWorld 81.9, WebArena-Verified 64.8, Vision2Web 62.9, MathVision 94.6, CharXiv (RQ) 90.2, and BabyVision 85.6. Several of these are in-house evaluations from the Qwen team: CoWorkBench covers long-horizon productivity domains, QwenSWEBench targets software engineering ability, and RecreationBench evaluates application recreation across desktop, mobile, and web.


Dense 27B vision-language model with hybrid Gated DeltaNet attention and 262K context

UD-Q8_K_XL

UD-Q5_K_XL

UD-Q4_K_XL

Qwen3.8 27B


MiniMax H3 is an omni-modal generative system from MiniMaxAI that produces video with native stereo audio from text, images, video, and audio inputs. It understands multimodal context as a single unified sequence rather than treating each input type separately, and predicts video and audio latents jointly in one forward pass so speech, sound effects, and music land in sync with the picture.

## Key Features

- **Native audio-video generation**: Video and 32 kHz stereo audio are modeled together, not stitched from separate systems
- **Multiple input modes**: Text-only, first-frame, last-frame, first-and-last-frame, and omni-reference conditioning
- **Flexible output shapes**: A wide range of aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16
- **24 FPS output** with durations from 4 to 15 seconds
- **Multilingual dialogue**: Stable support for Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish, with partial support for additional languages

## Architecture

The open-weights release is H3-Base, which generates 768p audio-video output. It encodes each modality with its own encoder or VAE, packs the encoded representations into one multimodal sequence, and applies three-dimensional multimodal rotary position embeddings across the temporal and two spatial dimensions before handing the sequence to the H3-Omni-Transformer.

The H3-Omni-Transformer is a 33B-parameter dense, single-stream transformer. Its attention and feed-forward layers carry no modality-specific structure; modality-specific parameters are confined to the input and output layers and to the AdaLN branches. Roughly 13B of the parameters sit in AdaLN-related branches whose modulation outputs can be precomputed and cached, so they do not need to be loaded for inference-only deployment.

Text and visual inputs are encoded by the H3-Encoder, which uses the full pretrained weights of Qwen3-VL-32B and passes hidden states from its 50th layer into the transformer. Visual inputs are additionally encoded by H3-VisualVAE, a temporally causal video autoencoder with 16x spatial compression, 4x temporal compression, and 24 latent channels. H3-AudioVAE compresses 32 kHz audio into latent tokens at a 40 Hz temporal rate, running the left and right channels independently through a shared encoder and decoder and recombining them for stereo.

## Model Variants

MiniMax H3 ships as two task-specific checkpoints:

- **H3-Base-FL2VA** handles first-and-last-frame mode. With no image input it runs text-to-video; with one image it generates from a first or last frame; with two images it generates between a first and last frame.
- **H3-Base-Ref2VA** handles omni-reference mode, accepting up to 9 images, up to 3 video clips, and up to 3 audio clips as references, with a maximum of 12 files across all input types and a total duration of 15 seconds.

Both checkpoints are CFG-distilled and released at BF16 precision.

## Use Cases

- Short-form video with dialogue, sound effects, and music generated together
- Character and product consistency across shots using omni-reference conditioning
- First-and-last-frame interpolation for controlled transitions
- Storyboard-to-motion previews from a single still frame
- Multilingual dialogue scenes without separate voice generation
- Creative experimentation with mixed image, video, and audio conditioning

## Prompting

Output quality depends heavily on how the multimodal context is described. MiniMaxAI ships a hosted preprocessing system, H3-Context-IR, that parses instructions, associates inputs across modalities, and serializes its understanding into a structured representation the base model consumes. That component is not part of the open-weights release, so detailed, explicit prompts that spell out the relationship between each reference input and the intended output matter more here than with single-modality video models.

## Integration

MiniMax H3 runs in ComfyUI through native nodes added in ComfyUI 0.30.0, with text-to-video, image-to-video, and reference-to-video workflow templates. Repackaged single-file weights for ComfyUI are published at [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3).

For more details about the model architecture and capabilities, see the [model page on Hugging Face](https://huggingface.co/MiniMaxAI/MiniMax-H3).


Audio-video foundation model generating up to 2K 24fps clips with natively synchronized stereo audio from text, image, or reference inputs

MiniMax H3


### Kimi K3

Kimi K3 is an open-weight, native multimodal agentic model from Moonshot AI and their most capable model to date. It is a 2.8T-parameter Mixture-of-Experts model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.

#### Key Features

- **New Architecture** — Built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), scaling MoE sparsity with a Stable LatentMoE framework that activates 16 of 896 experts, yielding roughly a 2.5x improvement in overall scaling efficiency over Kimi K2.
- **Long-Horizon Coding** — Operating with minimal human oversight, Kimi K3 sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools, from GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and chip design.
- **Agentic Knowledge Work** — Advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture.
- **Native Multimodality and Long Context** — Understands text, images, and video within the same model, and supports a 1-million-token context window.
- **Always-On Thinking** — Thinking is always enabled and returned as `reasoning_content`; thinking effort is selectable between low, high, and max. The model was trained in preserved thinking history mode, so prior assistant reasoning is passed back across multi-turn conversations and tool calls.
- **Open Frontier Weights** — The full model weights are openly released, making frontier-class multimodal agentic capability available to run on your own infrastructure.

#### Architecture

Kimi K3 is a Mixture-of-Experts model with 2.8T total parameters and 104B activated per token, spread across 93 layers (1 dense). Its attention stack composes 69 KDA layers with 24 Gated MLA layers over 96 attention heads and a 7168-dimensional attention hidden size. The Stable LatentMoE framework uses a 3584-dimensional latent MoE space with 3072 hidden dimensions per expert, selecting 16 of 896 experts per token alongside 2 shared experts. The model uses SiTU-GLU activations, a 160K vocabulary, and a 1,048,576-token context length. Quantization-aware training is applied from the SFT stage onward, giving native MXFP4 weights with MXFP8 activations for broad hardware compatibility.

On Vast, Kimi K3 is available as a full-precision vLLM flagship served from the original weights at [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3), and as single-node Unsloth dynamic 1-bit and 2-bit GGUF quantizations from [unsloth/Kimi-K3-GGUF](https://huggingface.co/unsloth/Kimi-K3-GGUF).

#### Benchmarks

On reasoning and knowledge, Kimi K3 scores 93.5 on GPQA Diamond, 74.7 on AA-LCR, 43.5 on HLE-Full (56.0 with tools), and 23.4 on CritPt.

On coding, it reaches 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE, 77.8 on ProgramBench, 72.9 on Kimi Code Bench 2.0, 67.5 on DeepSWE, 58.7 on SciCode, 48.3 on MLS-Bench-Lite, 42.0 on SWE-Marathon, and 36.6 on PostTrainBench.

On agentic tasks, it scores 95.0 on DeepSearchQA (F1), 94.6 on Harvey Lab-AA, 94.5 on MCPMark-Verified, 91.2 on BrowseComp, 84.8 on OSWorld-Verified, 84.2 on MCP-Atlas, 76.5 on Toolathlon-Verified, 76.2 on ResearchRubrics, 71.6 on CorpFin v2, 63.3 on OfficeQA Pro, 60.1 on SaaS-Bench, 58.3 on OSWorld 2.0, 54.4 on Finance Agent v2, 54.3 on JobBench, 44.2 on Legal Research Bench, 41.0 on APEX-Agents, 34.8 on SpreadsheetBench 2, 33.4 on tau-cubed Banking, 30.8 on AutomationBench, and 28.3 on Agents' Last Exam. It posts Elo ratings of 1686 on GDPval-AA v2 and 1548 on AA-Briefcase.

On vision, it scores 94.3 on MathVision (97.8 with python), 91.1 on OmniDocBench, 90.0 on Video-MME with subtitles, 85.7 on BabyVision with python, 84.8 on CharXiv reasoning questions (91.3 with python), 82.1 on MMVU, 81.6 on MMMU-Pro (83.4 with python), 58.5 on PerceptionBench, 51.0 on WorldVQA ForceAnswer, and 23.0 on ZeroBench pass@5 (41.0 with python).

#### Use Cases

- Long-horizon autonomous software engineering across large repositories
- Terminal and tool orchestration for build, test, and optimization loops
- Deep research producing interactive visualizations, dashboards, and widgets
- Document, chart, and screenshot understanding at scale
- Video understanding and motion design workflows
- Computer-use and browser agents operating over extended sessions
- Million-token context tasks spanning entire codebases or document sets


Moonshot AI's 2.8T-parameter multimodal MoE — full-precision vLLM flagship on 8xB300, plus single-node Unsloth 1-bit and 2-bit dynamic GGUF quants

UD-Q2_K_XL (Unsloth 2-bit)

UD-IQ1_M (Unsloth 1-bit)

UD-IQ1_S (Unsloth 1-bit)

Kimi K3


## Overview

Krea 2 Turbo is the distilled, few-step release checkpoint of the Krea 2 text-to-image model family from Krea.ai. It is a diffusion transformer with roughly 12 billion parameters that generates images from natural-language prompts. Turbo is post-trained and distilled from the Krea 2 Raw base checkpoint, so it produces high-quality images in far fewer sampling steps, making it well suited to fast, interactive generation.

## Key Features

- Few-step generation: distillation lets Turbo render detailed images in a handful of inference steps instead of the dozens a base diffusion model typically needs.
- Strong prompt following across photographic, illustrative, painterly, and stylized outputs.
- High-resolution synthesis suitable for detailed, print-quality imagery.
- Broad stylistic range, from photorealism and impressionist painting to retro, halftone, and low-poly aesthetics.

## Use Cases

Krea 2 Turbo is designed for creative, commercial, developer, and research workflows, including image generation, concepting, design exploration, visual production, and integration into applications and creative tools. Its low step count makes it a good fit for real-time or high-throughput generation where latency matters.

## Architecture

Krea 2 is a text-to-image diffusion transformer with about 12 billion parameters. The family ships two open-weight checkpoints: Krea 2 Raw, the base release prior to additional post-training, and Krea 2 Turbo, the post-trained and distilled checkpoint optimized for few-step sampling. Turbo pairs the diffusion transformer with a text encoder and VAE for end-to-end text-to-image synthesis.

## Model Family

This entry deploys Krea 2 Turbo. The base checkpoint it is distilled from is published as [Krea 2 Raw](https://huggingface.co/krea/Krea-2-Raw), and the full release is available at [Krea on Hugging Face](https://huggingface.co/krea).


Distilled few-step version of Krea 2, a 12B text-to-image diffusion transformer for fast high-quality image generation

Krea 2 Turbo


### GLM 5.2: 1M-Context Agentic Reasoning Model

GLM 5.2 is a 753B parameter Mixture-of-Experts model developed by Z.ai. It builds on GLM 5 with a sparse-attention design that supports a 1M-token context window, targeting long-horizon agentic tasks, large-codebase engineering, and advanced reasoning, with native English and Chinese support.

#### Key Features

- **Frontier Reasoning** - 99.2 on AIME 2026 and 91.2 on GPQA-Diamond
- **Agentic Software Engineering** - 62.1 on SWE-bench Pro for repository-level, multi-file tasks
- **Tool Use** - 40.5 on Humanity's Last Exam, rising to 54.7 with tool access
- **1M-Token Context** - Long-document analysis, large-repo navigation, and extended agentic trajectories in a single context
- **Interleaved Thinking** - Reasons before every response and tool call; defaults to thinking mode
- **Bilingual** - Native English and Chinese language support

#### Use Cases

- Software engineering, code generation, and multi-file repository-level tasks
- Multi-step agentic workflows with tool calling and web browsing
- Complex mathematical reasoning and competition-level problem solving
- Long-context document analysis, synthesis, and generation
- Terminal-based development, operations, and systems administration
- Research tasks requiring extended browsing and context management

#### Architecture and Design

GLM 5.2 is a 753B parameter Mixture-of-Experts model that uses a sparse attention mechanism with **IndexShare**, reusing the attention indexer across every four layers to reduce the cost of long-context inference while preserving capacity across its 1M-token window. The design extends the GLM 5 architecture toward longer context and more reliable agentic behavior.

#### Training and Inference

GLM 5.2 builds on the GLM 5 foundation with refreshed post-training for stronger agentic coding and tool use. It defaults to thinking mode, reasoning before each response and tool call, and Z.ai recommends a temperature of 1.0 with top-p 0.95 for general reasoning tasks.

Deploy GLM 5.2 on Vast.ai for frontier-class agentic reasoning, coding, and long-context capabilities with flexible GPU infrastructure.


753B MoE model with 1M-token context for agentic reasoning, coding, and tool use

Q4_K_M (Unsloth)

GLM 5.2


### EmbeddingGemma 2: Multimodal Embeddings for Text, Images, Video and Audio

EmbeddingGemma 2 is an open multimodal embedding model from Google DeepMind. It maps text (including code), images, video, and audio, and combinations of them, into a single shared 768-dimensional vector space. A text query can therefore be compared directly against an image, a video clip, or an audio recording. The model pairs a compact text backbone with modular vision and audio encoders and builds on the architecture of Gemma 4.

#### Key Features

- Native multimodality: text, images, video, and audio share one embedding space
- Multilingual and code-aware: understands more than 100 languages, with a reported 14% gain on code tasks over the first EmbeddingGemma
- 8,192-token context shared across modalities, enough for minutes of audio or video in one input
- Matryoshka Representation Learning: the 768-dimensional vector can be truncated to 512, 256, or 128 dimensions for up to 6x lower storage
- Task-steered embeddings through short text instruction prefixes
- Interleaved inputs: one embedding can represent text mixed with several images, video, and audio

#### Task Instruction Prefixes

EmbeddingGemma 2 is trained with short task prefixes prepended to text inputs. Using the right prefix improves quality; omitting it still works but reduces precision. Images, video, and audio take no prefix.

For retrieval-style tasks, prefix queries and documents differently:

- Web or document search: `task: search result | query: {query}`
- Question answering: `task: question answering | query: {question}`
- Fact checking: `task: fact checking | query: {claim}`
- Code search: `task: code retrieval | query: {query}`
- Documents for any of these: `title: {title} | text: {content}`, using `title: none` when there is no title

For symmetric tasks, apply the same prefix to every input being compared:

- Classification: `task: classification | query: {content}`
- Clustering: `task: clustering | query: {content}`
- Semantic similarity: `task: sentence similarity | query: {content}`

#### Matryoshka Dimensions

The supported output sizes are 768, 512, 256, and 128. This template returns the full 768-dimensional vector. To store a shorter one, keep its leading dimensions and L2-normalize the result again before computing cosine similarity; the model card warns that skipping re-normalization silently degrades ranking quality. Queries and documents must share a dimension. Quality stays close to lossless down to 256 dimensions, while 128 dimensions degrades multimodal quality substantially.

#### Architecture

EmbeddingGemma 2 has a 24-layer text backbone with a 5:1 ratio of local sliding-window to global attention layers, grouped-query attention, and a 262,144-token vocabulary. Token representations are mean-pooled and projected from 512 to 768 dimensions. Separate vision and audio encoders feed the same backbone. Each image costs about 280 tokens by default, video is sampled at one frame per second at about 140 tokens per frame, and audio costs 25 tokens per second, all drawn from the shared 8,192-token budget. The model should run in bfloat16 or float32; the model card warns that float16 produces NaN or degraded embeddings.

#### Use Cases

- Semantic search and RAG over text, code, images, video, and audio
- Code search from natural-language queries
- Spoken-query search over audio archives
- Classification, such as sentiment analysis, content moderation, audio event detection, or image categorization
- Clustering and near-duplicate detection across modalities
- Evidence retrieval for fact-checking and source attribution

#### Benchmarks

Google reports a multilingual MTEB (v2) mean of 61.36 at 768 dimensions, slightly above the first EmbeddingGemma's 61.15, and a code MTEB (v1) score of 78.68 against its predecessor's 68.76. On image tasks the model scores 64.64 on MIEB lite and 57.28 on the MMEB v2 image split, with 67.84 on visual document retrieval. It reaches 50.67 on MMEB v2 video, 69.54 on MSEB audio retrieval, and 49.39 on MAEB. Truncating to 256 dimensions keeps multilingual MTEB at 60.41, while 128 dimensions drops it to 57.89 and costs more on multimodal benchmarks.

Learn more on the [EmbeddingGemma 2 model card](https://huggingface.co/google/embeddinggemma-2).


Google's multimodal embedding model mapping text, code, images, video, and audio into one 768-dimensional vector space, served as an OpenAI-compatible embeddings API

EmbeddingGemma 2


### Flint: Open-Weight MoE Model for Coding Agents

Flint is Recursive's mixture-of-experts language model for coding agents. It is designed to investigate codebases, fix bugs, implement features, and write tests, and it supports reasoning and multi-turn tool use. Flint was trained from scratch through pretraining, mid-training, supervised fine-tuning, and reinforcement learning.

The model is published at [RecursiveSI/flint](https://huggingface.co/RecursiveSI/flint).

#### Key Features

- **Built for coding agents** - Trained on text, code, instruction following, and agentic tasks, aimed at agents that work through a codebase over many steps.
- **Small active footprint** - About 2.5B of its 39.8B parameters are active per token, so decoding is fast for a model of its total size.
- **Reasoning** - Thinks before answering; the reasoning is returned separately from the final answer.
- **Tool calling** - OpenAI-compatible tool calls for multi-turn agent workflows.
- **Long context** - A 262,144-token context window shared between prompt and output.

#### Architecture

Flint is a decoder-only mixture-of-experts transformer in which each token is routed to a small set of experts plus an always-active shared expert. Most layers use local sliding-window attention, and only a minority attend to the full context, so long contexts stay cheap in memory.

#### Benchmarks

In Recursive's release evaluations Flint scores 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified. Agentic scores are the average pass@1 over at least five attempts, with Terminal-Bench run through Terminus 2 and SWE-bench through mini-swe-agent.

#### Use Cases

- Coding agents that read a repository, then edit and test code
- Bug fixing and feature implementation
- Writing tests for existing code
- Tool-using assistants built on OpenAI-compatible function calling

#### Sampling

The publisher recommends a temperature of 1.0 with top-p 0.95 and top-k 20, and an output budget of at least 32,768 tokens on hard problems, since reasoning counts toward the output limit. Keep returned reasoning content in the conversation history across turns.


Recursive's 40B-parameter coding MoE with about 2.5B active parameters per token and a 262K-token context, built for coding agents

UD-Q3_K_XL

Q8_0

Flint


### GLM-OCR: Compact Document OCR and Information Extraction

GLM-OCR is a multimodal OCR model from Z.ai for complex document understanding, built on the GLM-V encoder-decoder architecture ([zai-org/GLM-OCR](https://huggingface.co/zai-org/GLM-OCR)). It reads page images and returns their text, tables, and formulas, and it can fill a caller-supplied JSON schema with structured fields pulled from a document.

#### Key Features

GLM-OCR scores 94.62 on OmniDocBench V1.5, and the publisher reports strong results across formula recognition, table recognition, and information extraction benchmarks.

The model is tuned for real business documents rather than clean scans. It keeps its accuracy on complex tables, code-heavy pages, seals and stamps, and other difficult layouts.

Training introduces a Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning, which the publisher credits for better training efficiency, recognition accuracy, and generalization.

The model is small and fast. In the publisher's single-replica, single-concurrency speed test it parsed 1.86 PDF pages per second and 0.67 images per second when exporting Markdown.

#### Prompting

GLM-OCR works with two kinds of prompts. For document parsing, send the image with one of three task prompts: `Text Recognition:`, `Formula Recognition:`, or `Table Recognition:`. For information extraction, send the image with a JSON template whose keys name the fields to fill (for example ID number, name, date of birth, and address); the model's output follows that schema.

#### Architecture

GLM-OCR pairs a CogViT visual encoder, pre-trained on large-scale image-text data, with a lightweight cross-modal connector that downsamples visual tokens efficiently and a GLM-0.5B language decoder. For full-page document parsing, the publisher's SDK runs a two-stage pipeline: PP-DocLayout-V3 first detects the page layout, then GLM-OCR recognizes each region in parallel and the SDK assembles structured output. The served model handles the recognition stage and direct information extraction.

#### Use Cases

- Converting scanned pages and PDFs into Markdown for search, RAG, and LLM training data
- Table extraction with structure preserved
- Formula recognition for scientific and technical documents
- Text extraction from code screenshots
- Structured field extraction from IDs, forms, receipts, and invoices into fixed JSON schemas
- Multilingual OCR across Chinese, English, French, Spanish, Russian, German, Japanese, and Korean


Compact 0.9B OCR vision-language model for document parsing, tables, formulas and structured information extraction

GLM-OCR


### MiniCPM-V 4.7 35B A3B: Efficient Vision-Language MoE

MiniCPM-V 4.7 is a release in OpenBMB's MiniCPM-V series of multimodal models, and the 35B A3B checkpoint is the first Mixture-of-Experts model in the series. It has about 35B total parameters and activates about 3B per token, and it handles single-image, multi-image and video understanding. A dense 1B sibling, MiniCPM-V 4.7 1B, uses the same vision encoder and hybrid-attention design without the expert layers, and is offered here as a variant.

The vision encoder applies the intra-ViT early compression technique from LLaVA-UHD v4, which compresses visual tokens inside the encoder and cuts the work done by later encoder layers. The model also supports mixed 4x and 16x visual token compression, so each request can trade visual detail against inference cost.

#### Key Features

- **Sparse Mixture-of-Experts** - 256 experts with about 3B of the 35B parameters active per token
- **Hybrid Attention Backbone** - Alternates linear-attention layers with full attention, one full-attention layer in every four
- **Image, Multi-Image and Video Input** - Accepts images and video clips alongside text in the same conversation
- **Adjustable Visual Token Compression** - Mixed 4x and 16x modes balance detail against speed
- **Optional Thinking Mode** - Answers directly by default, with a thinking mode that can be turned on per request
- **Tool Calling** - Emits function calls that the server parses into standard OpenAI-style tool calls
- **Native 262K Context** - Handles 262,144 tokens of context

#### Use Cases

- Visual question answering over photos, screenshots and diagrams
- Document, chart and on-screen text understanding
- Video description and timeline summarization
- Multi-image comparison and reasoning
- Agents that combine visual input with tool calls

#### Benchmark Performance

OpenBMB reports that, in Instruct mode, the 35B A3B model combines a high overall score with short inference time across its vision-language evaluations. The 1B variant is reported to improve on the foundation capabilities of MiniCPM-V 4.6.

#### Thinking Mode

Thinking mode is off by default, so requests get a direct answer. To turn it on for a request, send `"chat_template_kwargs": {"enable_thinking": true}` in the chat completions body.

#### Recommended Sampling

For Instruct mode the publisher recommends temperature 0.7, top_p 0.95, top_k 100, min_p 0.0, presence_penalty 0 and repetition_penalty 1.03.

Deploy MiniCPM-V 4.7 35B A3B on Vast.ai with vLLM for image and video understanding at MoE efficiency.

More details are on the [MiniCPM-V 4.7 35B A3B model card](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B).


First MoE model in the MiniCPM-V series, with 3B active parameters for image, multi-image and video understanding

MiniCPM-V 4.7 35B A3B


### Qwen3.5 9B: Compact Vision-Language Reasoning Model

Qwen3.5 9B is the largest of the small dense models in Alibaba's Qwen3.5 family. It is a native vision-language model: text, images, and video are trained together from the start, so one checkpoint handles chat, coding, document reading, and visual reasoning. This page also offers the 4B, 2B, and 0.8B sizes, which share the same architecture and context window for lighter hardware.

#### Key Features

-   **Unified vision and language** - early-fusion training on multimodal tokens, so images and video are first-class inputs rather than an add-on adapter
-   **Thinking by default** - the 9B and 4B open each answer with a reasoning block, returned separately from the final answer; the chat template's `enable_thinking` switch turns it off. The 2B and 0.8B default to direct answers
-   **Tool calling** - structured function calls returned as standard OpenAI-style tool calls, ready for agent frameworks
-   **Long context** - 262,144 tokens natively, with the publisher documenting YaRN scaling up to about one million
-   **201 languages and dialects** - broad multilingual coverage for global deployments
-   **Four sizes** - 9B, 4B, 2B, and 0.8B, with the same architecture and 262K context

#### Architecture

Qwen3.5 9B has 32 layers arranged in eight repeating blocks of three Gated DeltaNet (linear attention) layers followed by one gated full-attention layer. The linear layers keep a fixed-size state instead of a growing key-value cache, so memory rises slowly as context gets longer and long prompts stay affordable. A vision encoder feeds image and video tokens into the same backbone. The model was also trained with multi-token prediction.

#### Benchmarks

On Alibaba's published results, the 9B scores 82.5 on MMLU-Pro, 81.7 on GPQA Diamond, 91.5 on IFEval, and 83.2 on HMMT February 2025. For agents it reaches 66.1 on BFCL-V4 and 79.1 on TAU2-Bench. On vision it scores 78.4 on MMMU, 70.1 on MMMU-Pro, 78.9 on MathVision, 89.2 on OCRBench, 87.7 on OmniDocBench 1.5, and 78.9 on VideoMMMU. The 4B scores 79.1 on MMLU-Pro, 76.2 on GPQA Diamond, 77.6 on MMMU, and 79.9 on TAU2-Bench.

#### Use Cases

-   Visual question answering, chart reading, and document OCR
-   Coding assistants and tool-calling agents
-   GUI and screen understanding for computer-use agents
-   Video understanding and summarization
-   Multilingual chat and translation
-   Edge-sized deployments with the 2B and 0.8B variants

Qwen3.5 9B gives teams a capable multimodal reasoning model in a compact size, with smaller sizes on the same page when cost or latency matters most.


Compact dense vision-language model with a 262K context, in 9B, 4B, 2B, and 0.8B sizes

0.8B

9B Q8_0

9B UD-Q4_K_XL

Qwen3.5 9B


### NVIDIA Nemotron 3.5 Bolt 74B A9B: Efficient Vision-Language Agent Model

NVIDIA Nemotron 3.5 Bolt 74B A9B is a vision-language model from NVIDIA built to run everyday agentic tasks: tool calling, web browsing and terminal use. It reads text, images and video and answers in text. NVIDIA released it on Hugging Face on October 15, 2026.

It ships in two forms. The [BF16 release](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16) holds the full-precision weights and is the reference checkpoint for customization and post-training. The 24GB release uses a native 2-bit quantization; NVIDIA recommends it for latency- and throughput-optimized inference, and publishes it both as a [GGUF for llama.cpp](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF) and as [safetensors for vLLM](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB).

#### Architecture

Bolt is a hybrid that interleaves Mamba-2 state-space layers, Mixture-of-Experts layers and a small number of attention layers, using NVIDIA's LatentMoE design. Images and video frames go through a C-RADIO v4-H vision encoder, connected to the language model by an MLP projector. The model supports up to 256K tokens of context, shared by the prompt, visual inputs, conversation history, tool results and the generated response.

It was compressed from [Nemotron 3.5 Super VL 120B-A12B](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16) with Star Elastic, NVIDIA's router-guided architecture search and compression method. A budget-conditioned router chose the number of experts and the expert width separately for each layer, while the number of experts used per token was fixed at 9, down from 22 in the parent. Knowledge recovery ran for 120B tokens, first at 8K context and then at 128K.

#### Training

After compression, NVIDIA re-attached the parent model's vision tower and retrained the vision adapter on about 6.6M image-text samples covering captioning, visual question answering, counting, documents and OCR. A vision-heavy supervised fine-tuning stage followed, targeting GUI grounding and web navigation alongside text reasoning, coding and tool-use data, with checkpoint merging to keep text and vision skills balanced. Reinforcement learning then trained the model across agentic environments for MCP tool use and software engineering, with an RLHF stage that also penalized excess reasoning for token efficiency.

For the 24GB release, the experts were quantized to 2-bit (IQ2_XXS) with quantization-aware distillation to recover accuracy. Mamba layers are kept in NVFP4 and the embedding and output layers in 8-bit.

#### Key Features

- Text, image and video input, with video clips of up to two minutes
- Reasoning on or off per request through the chat template, plus a medium-effort reasoning mode
- Native tool calling for agent harnesses
- Up to 256K tokens of context
- English and programming languages as primary, with Spanish, French, German, Italian and Japanese also supported
- Post-training data through May 2026

#### Benchmarks

NVIDIA reports that the 24GB release scores 86.43 on GPQA Diamond, 50.98 on Terminal-Bench 2.1, 66.75 on BrowseComp, 52.20 on MCP-Atlas, 50.86 on WebArena with vision, 64.06 on AA-LCR, 41.02 on SciCode, 25.30 on HLE without tools, 28.47 on GDPVal-AA-V2 and 87.82 on PinchBench.

The full-precision BF16 release scores a few points higher on most of the same tests, for example 88.83 on GPQA Diamond, 53.93 on Terminal-Bench 2.1, 70.00 on BrowseComp, 55.60 on MCP-Atlas and 90.35 on PinchBench.

#### Use Cases

- Local and self-hosted agents that call tools, browse the web and work in a terminal
- Visual tasks such as screen and GUI understanding, documents, charts and web navigation
- Video understanding and summarization
- Long-document analysis and retrieval-augmented generation
- Post-training, domain adaptation and building your own quantized variants from the BF16 weights

#### Sampling

NVIDIA recommends a temperature of 1.0 and top-p of 0.95.


Hybrid Mamba-Transformer MoE vision-language model with 9B active params, built for local agentic work with tool calling and reasoning control

24GB

NVIDIA Nemotron 3.5 Bolt 74B A9B


### NVIDIA Nemotron 3.5 Super VL 120B A12B: Long-Context Vision-Language Reasoning

NVIDIA Nemotron 3.5 Super VL is a vision-language model from NVIDIA built for reasoning, coding, tool use, document understanding and video understanding. It reads text, images and video and answers in text. The model has 120 billion total parameters with 12 billion active per token, supports contexts of up to 1 million tokens, and was released on Hugging Face on October 15, 2026.

Nemotron is NVIDIA's family of open models, with open weights, open training data and open recipes, aimed at building specialized AI agents efficiently. NVIDIA publishes a full-precision [BF16 release](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16) and an [NVFP4 release](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4) refined with quantization-aware distillation from the BF16 weights.

#### Architecture

The language backbone interleaves Mamba-2 state-space layers, Mixture-of-Experts layers and a small number of attention layers. Of its 88 layers, 40 are Mamba-2, 40 are MoE and 8 are attention, which keeps the active parameter count low and lets the model hold very long contexts without the memory growth of a pure attention stack. Each MoE layer has 512 routed experts plus a shared expert.

A C-RADIO vision encoder and a vision-language connector turn images, document pages and sampled video frames into representations the language backbone processes. A shared Multi-Token Prediction head predicts future tokens for native speculative decoding.

#### Key Features

- Image and video understanding on top of a strong text reasoning model, with documents supplied as page images alongside text
- Video input as MP4 clips of up to two minutes, sampled into frames
- Configurable reasoning: off, on (medium) and on, selected through the chat template; medium effort produces fewer reasoning tokens at some cost in accuracy
- Native tool calling, with reasoning and tool-call output parsed into separate response fields
- Context window of up to 1 million tokens, shared by the prompt, visual inputs, conversation history, tool results and the response
- Multilingual training data covering English, 19 other spoken languages and 43 programming languages

#### Training

The language backbone starts from Nemotron 3 Super Base. NVIDIA first trained the shared MTP head with the backbone frozen, then distilled from Nemotron 3 Ultra Base, and ran continued pre-training that includes an extension to 1 million tokens of context and software-engineering mid-training on repositories, issue and pull-request pairs, build logs and agent trajectories. Vision-language training then added the C-RADIO encoder and connector, followed by supervised fine-tuning on packed sequences of up to 524,288 tokens, reinforcement learning with verifiable rewards and human feedback, and multi-teacher on-policy distillation. Pre-training data runs through December 2025.

#### Benchmarks

NVIDIA collected all of its results with NeMo Gym under consistent settings.

On reasoning and knowledge the model scores 89.52 on GPQA-Diamond, 31.79 on text-only Humanity's Last Exam and 88.36 on the APEX shortlist. On coding it reaches 88.22 on LiveCodeBench v6, 71.10 on SWE-bench Verified, 51.78 on SWE-bench Pro and 60.67 on Terminal-Bench 2.1. On agentic and long-context work it scores 78 on BrowseComp, 76.50 on tau2-bench airline, 89.71 on GraphWalks and 68.62 on AA-LCR. Its hallucination rate on AA-Omniscience is 30.0 percent.

On vision it scores 85.60 on V*, 80.82 on CharXiv reasoning, 78.16 on MathVision, 65.08 on OCR-Reasoning, 62.82 on MMLongBench-Doc and 51.14 on OSWorld-Verified.

#### Use Cases

- Agents that read screenshots, documents and charts and act through tool calls
- Coding assistants and software-engineering agents
- Document question answering over scanned pages, forms and long reports
- Video summarization and question answering on short clips
- Research and search workflows that combine reasoning with tools

#### Sampling

NVIDIA recommends a temperature of 1.0 and top-p of 0.95 for text, and a temperature of 0.6 and top-p of 0.95 for vision tasks. Reasoning tokens and the final answer share the max_tokens budget, so leave room for both.


Hybrid Mamba-attention MoE vision-language model with 12B active params, image and video input, reasoning control and tool calling

NVFP4

NVIDIA Nemotron 3.5 Super VL 120B A12B


### Seed 2.0 Pro

Seed 2.0 Pro is ByteDance Seed's flagship large language model, served on Vast as a Model Endpoint through an OpenAI-compatible API.


ByteDance Seed 2.0 Pro, served as a Vast Model Endpoint with an OpenAI-compatible API

Seed 2.0 Pro


### Seedream 5.0 Pro

Seedream 5.0 Pro is ByteDance Seed's image generation model, served on Vast as a Model Endpoint through an OpenAI-compatible images API.


ByteDance Seedream 5.0 Pro image generation, served as a Vast Model Endpoint

Seedream 5.0 Pro


### K2 Horizon MoVA 36B-A4B

K2 Horizon MoVA 36B-A4B is the sparse member of IFM's [K2 Horizon](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B) family. It is a Mixture-of-Experts model that also uses Mixture-of-Values attention (MoVA), storing 36B parameters and running about 4B per token. IFM releases it fully open: the final checkpoint, the intermediate training checkpoints, the training data and recipe, and the training code are all public.

#### Key Features

- **Sparse and efficient** - routed experts in the feed-forward layers plus routed value experts inside attention, so only a small slice of the model runs for each token.
- **512K-token context** - native 524,288-token context from the midtraining stages onward.
- **Selectable reasoning depth** - thinking effort can be set per request to high (the default), medium, or low.
- **Tool calling** - supports JSON, XML, and typed-XML tool-call formats, chosen per request.
- **Intermediate checkpoints** - checkpoints from every training stage are released, so capability changes can be studied across training rather than at a single point.
- **Fully open** - training data, recipe, and code are published alongside the weights.

#### Multi-Turn Conversations

The server returns the model's thinking in the `reasoning` field, separate from the answer in `content`. When you send a conversation back, every earlier assistant turn must carry that thinking in `reasoning` (or `reasoning_content`), including turns that only make tool calls. An assistant turn without it is rejected with HTTP 400 "Assistant message is missing a thinking field".

```json
{
  "model": "IFM/K2-Horizon-MoVA-36B-A4B",
  "messages": [
    {"role": "user", "content": "What is 17 x 23?"},
    {"role": "assistant", "content": "17 x 23 = 391.", "reasoning": "17 x 20 = 340, 17 x 3 = 51, 340 + 51 = 391."},
    {"role": "user", "content": "Now divide that by 7."}
  ]
}
```

If your client does not keep the reasoning, send `"reasoning": ""` on each earlier assistant turn. An empty string is accepted; the model then sees that turn's answer with an empty thinking block.

#### Training

The model was pretrained on 22.9T tokens at an 8K sequence length. Four midtraining stages then extended the context to 32K, 128K, and finally 512K, with the last stage shifting the data mix toward agentic and reasoning data. Two phases of supervised fine-tuning at 512K followed, the second on a high-quality subset of the first with learning-rate decay.

#### Benchmark Highlights

On its published evaluations, IFM reports 26.8 on tau3-Banking for agentic tool use, 58.6 on Terminal-Bench 2.1 for agentic terminal use, and 38.9 on SciCode. On scientific reasoning it reaches 80.8 on GPQA Diamond and 25.2 on Humanity's Last Exam without tools. For long-context reasoning it scores 66.3 on AA-LCR, and it posts a 69.2 non-hallucination rate on AA-Omniscience.

#### Use Cases

- Agentic workflows and tool use
- Coding and terminal automation
- Scientific and multi-step reasoning
- Long-document analysis over very long contexts
- Research on open models and on how capabilities develop during training

Deploy K2 Horizon MoVA 36B-A4B on Vast.ai to run a fully open, efficient Mixture-of-Experts model with vLLM.


Fully open MoE with Mixture-of-Values attention, 4B active parameters and 512K context

K2 Horizon MoVA 36B-A4B


### MiMo V2.6 Pro RL

MiMo-V2.6-Pro-RL is the flagship checkpoint of Xiaomi's MiMo-V2.6 series, built around a single idea: scale reinforcement learning toward self-improvement. Rather than training separate specialists per domain, the model is trained in one mixed RL run spanning coding, general agent work, visual tasks, and cybersecurity, so that capabilities reinforce one another and learned strategies transfer to harnesses the model never saw in training.

It is natively omnimodal. Text, image, video, and audio are handled in one model, with a context window of 1M tokens intended for long repositories, extended tool traces, and multi-session agent runs.

#### Key Features

**Native omnimodal input.** Text, image, video, and audio are first-class inputs, handled by the same model.

**One million token context.** Long enough to hold entire repositories, long tool-call histories, and agent sessions that span many turns.

**You only RL once.** A single mixed reinforcement learning run covers coding, general agents, visual reasoning, and cybersecurity. Tasks and multiple harnesses are mixed into the same batch, which is what allows strategies to transfer across domains.

**Scaled RL compute.** Fully asynchronous Group Relative Policy Optimization over very large batches, at 1,568 prompts by 16 rollouts per step and billions of tokens per update.

**Groupwise agentic grading.** Binary pass or fail cannot rank solutions that all pass, so the reward signal itself is scaled. Groupwise Reward Synthesis builds task-specific rubrics offline from contrasting rollouts and fuses rubric quality with test outcomes. Groupwise Advantage Redistribution ranks passing trajectories online and shifts advantage toward higher-quality solutions. Because the model is judged against its own samples, this closes a self-improvement loop that favors shorter paths and fewer tokens per task.

**Aligned RL.** Training cold-starts from self-correction, where the model reflects on and rewrites its own misaligned turns into grounded next steps. Environment hardening, adversarial screening, and verifier cross-checks run throughout to keep the loop honest against reward hacking.

**Multi-prefix multi-teacher on-policy distillation.** After mixed RL, MOPD2 combines autonomous student rollouts with prefix-conditioned single-turn rollouts, reusing histories from teacher trajectories and demonstrations so that decision points can be trained without regenerating everything that preceded them. This extends the model into tasks that are hard to verify automatically.

**Reasoning and tool calling.** The model emits separate reasoning and answer channels and produces structured tool calls, with reasoning switchable per request.

#### Architecture

MiMo-V2.6-Pro-RL is a sparse Mixture-of-Experts model with 1.02 trillion total parameters, of which 42 billion are active on any given token. The backbone is 70 layers deep. The first block uses global attention with a dense feed-forward network; the remaining blocks interleave local sliding-window attention and global attention, and all of them use sparse MoE feed-forward networks with no shared experts. Ten layers use global attention and sixty use a sliding window of 128 tokens, which is what keeps a one-million-token context tractable. There are 384 routed experts with 8 activated per token.

Vision is handled by MiMo ViT, a 681M-parameter encoder of 28 layers, 24 sliding-window and 4 global, operating on 2 by 16 by 16 patches with 2 by 2 spatial merging. Audio runs through a 308M-parameter AudioTokenizer encoder using 20 residual vector-quantization codebooks, paired with a 127M-parameter audio patch encoder that groups four frames per patch.

The checkpoint also ships a speculative decoder: a five-layer sliding-window multi-token-prediction drafter that proposes seven subsequent tokens per forward pass for parallel verification.

#### Benchmarks

Xiaomi reports evaluation across four families of agentic task, and the headline result is the size of the jump over the previous generation rather than any single score.

On code agent work, MiMo-V2.6 Pro reaches 71.9 on DeepSWE v1.1 and 63.2 on MiMo Code Bench, against 19.0 and 40.4 respectively for MiMo-V2.5 Pro. On general agent benchmarks it scores 53.1 on AutomationBench v1.0.6, 76.9 on Toolathlon-Verified, 89.9 on Terminal Bench 2.1, 34.9 on the harder Terminal Bench 4.0, 82.0 on OSWorld-Verified, and 62.0 on JobBench, with an Agents' Last Exam score of 31.6. Cybersecurity shows the largest generational gap: 94.0 on CyberGym and 80.2 on MiMo Cyber Bench, where the previous generation scored 40.0 and 0.0. On visual agent work it records 72.3 on MiMo VisualCoding.

#### Use Cases

- Long-horizon coding agents working across whole repositories and extended tool traces
- General computer-use and automation agents that must operate unfamiliar harnesses
- Multimodal document, video, and audio understanding in a single pipeline
- Security research workflows, including vulnerability discovery and analysis
- Visual coding and UI reasoning from screenshots or recordings
- Multi-session assistants that need to carry very large contexts between turns

Model weights and the full technical report are published at [XiaomiMiMo/MiMo-V2.6-Pro-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL). A smaller sibling in the same series is available as [XiaomiMiMo/MiMo-V2.6-Flash-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL).


Xiaomi's flagship MoE reasoning model - 1.02T total / 42B active params, trained with large-scale mixed RL for coding, agents and cybersecurity, with a 1M-token context window

MiMo V2.6 Pro RL


### Inkling-Small

Inkling-Small is a general-purpose multimodal model from Thinking Machines. It accepts text, image and audio inputs and generates text. It is a sparse Mixture-of-Experts model with 276B total parameters and 12B activated per token, built for developers making agentic and tool-use systems, coding assistants, chatbots and retrieval-augmented generation applications.

#### Overview

Inkling-Small is a 42-layer decoder-only transformer with a sparse Mixture-of-Experts feed-forward backbone. Each token is routed to 6 of 256 experts, plus 2 shared experts that are active on every token. Attention is a hybrid of local and global layers. The model is natively multimodal: images are encoded with a hierarchical patch encoder and audio through discrete token encoding, and all modalities are projected into the same model. This template serves the official NVFP4 checkpoint published by Thinking Machines.

#### Key Features

- Text, image and audio input with text output
- Sparse MoE routing that activates 12B of 276B parameters per token
- Hybrid local and global attention with a context window of up to 1M tokens
- Reasoning and tool calling for agentic workflows
- English, with general multilingual capability and broad coverage of programming languages

#### Inputs

Images can be any pixel-based input; Thinking Machines recommends keeping each dimension between 40 and 4096 pixels. Audio is WAV at 16 kHz, ideally under two minutes per clip.

#### Benchmarks

Thinking Machines reports Inkling-Small at 80.2% on SWE-bench Verified (bash-only harness), 55.9% on SWE-Bench Pro (public), 64.7% on Terminal-Bench 2.1, 79.6% on MCP Atlas (public) and 77.4% on BrowseComp with context management. On reasoning it scores 89.5% on GPQA Diamond, 95.5% on AIME 2026, 90.2% on HMMT February 2026 and 31.6% on Humanity's Last Exam (text only, 47.8% with tools). On instruction following it reaches 82.2% on IFBench. Multimodal results include 74.0% on MMMU Pro, 77.0% on MMAU and 90.1% on VoiceBench. The model card compares these results against open-weight models including Qwen3.5 397B-A17B, MiMo V2.5, MiniMax M2.7 and DeepSeek V4 Flash.

#### Use Cases

- Coding assistants and autonomous software engineering agents
- Tool-calling and browsing agents
- Chat and retrieval-augmented generation over long documents
- Understanding images, charts and screenshots
- Answering questions about speech and other audio

For full model details, evaluation notes and safety information, see the model card on [Hugging Face](https://huggingface.co/thinkingmachines/Inkling-Small).


Thinking Machines' 276B MoE (12B active) that takes text, images and audio, with a 1M-token context for coding, agents and reasoning

Inkling-Small


### MiMo V2.6 Flash RL: Agentic Mixture-of-Experts with a 1M-Token Context

MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of Xiaomi's MiMo-V2.6 series, a sparse Mixture-of-Experts model with 309B total parameters and 15B activated per token. It is built for long-horizon agent work: coding across large repositories, terminal and computer-use agents, and cybersecurity tasks, with a 1-million-token context window. The series is trained around one idea, scaling reinforcement learning toward self-improvement, by growing RL compute, environment diversity, and grader compute together.

#### Key Features

-   **Long horizon** - a 1M-token context window for long repositories, extended tool traces, and multi-session agent runs
-   **Native omnimodal design** - the model carries its own vision and audio encoders alongside the text backbone, covering text, image, video, and audio in one model
-   **You Only RL Once** - one mixed RL run across coding, general agents, visual work, and cybersecurity, rather than separate per-domain runs, so strategies transfer to harnesses never seen in training
-   **Groupwise agentic grading** - an agentic grader compares rollouts within each group to rank passing solutions, steering the model toward shorter paths and fewer tokens per task
-   **Thinking mode** - the model reasons before it answers, with the reasoning returned separately from the final answer; setting the chat template's `enable_thinking` option to false turns it off per request
-   **Tool calling** - structured tool calls for agent frameworks

#### Architecture

The language backbone has 48 layers. The first uses global attention with a dense feed-forward network; the rest interleave sliding-window and global attention, 39 sliding-window layers with a 128-token window and 9 global layers, and all of them use sparse Mixture-of-Experts feed-forward networks that route each token to 8 of 256 experts. Because only the 9 global layers keep a key-value cache that grows with context length, memory at long context grows far more slowly than in a conventional transformer of similar depth. The expert weights are stored in MXFP4 with a BF16 router. A 681M-parameter vision encoder and a two-stage audio encoder feed the same backbone, and the checkpoint also ships a 5-layer speculative drafter that predicts seven tokens per forward pass.

#### Benchmarks

Xiaomi reports the Flash checkpoint close behind the larger MiMo-V2.6 Pro on most evaluations while activating far fewer parameters per token. On code agent tasks it scores 67.9 on DeepSWE v1.1, 61.2 on MiMo Code Bench, and 26.0 on ProgramBench. On general agent tasks it reaches 87.6 on Terminal Bench 2.1, 80.8 on OSWorld-Verified, 73.6 on Toolathlon-Verified, 61.2 on JobBench, and 52.3 on AutomationBench v1.0.6. On cybersecurity it posts 95.1 on CyberGym and 77.2 on MiMo Cyber Bench, and on visual agent work it scores 71.5 on MiMo VisualCoding. The previous MiMo-V2.5 Pro scored 19.0 on DeepSWE v1.1 and 65.2 on Terminal Bench 2.1.

#### Use Cases

-   Long-horizon coding agents that navigate large repositories, run terminal tools, and sustain multi-step engineering sessions
-   Computer-use and tool-calling agents for desktop, office, and job workflows
-   Security research and vulnerability analysis
-   Long-context analysis where full repositories, tool traces, or multi-session histories must stay in the prompt

Full model details are available on the [MiMo-V2.6-Flash-RL model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL). The larger sibling is [MiMo-V2.6-Pro-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL).


Xiaomi's 309B-parameter Mixture-of-Experts model (15B active) for agentic coding, tool use, and cybersecurity, with a 1M-token context window

MiMo V2.6 Flash RL


### Ornith 1.5 35B A3B: Self-Improving Agentic Coding MoE

Ornith 1.5 35B A3B is the mid-size Mixture-of-Experts member of the Ornith-1.5 family from Ornith. It has about 35B total parameters and activates about 3B per token. Ornith-1.5 builds on Ornith-1.0, which was developed on top of Qwen3.5 and Gemma 4 with additional continued pretraining, mid-training and post-training. This checkpoint uses the Qwen3.5 MoE architecture.

Ornith-1.5 is trained through end-to-end self-improvement. Rather than relying on a fixed set of human-curated tasks and hand-designed harnesses, the training loop generates new tasks, builds the scaffolds to solve them, discovers effective strategies, and improves the policy through reinforcement learning.

#### Key Features

- **Agentic Coding** - Built for terminal-based coding agents and tool-calling workflows
- **Sparse Mixture-of-Experts** - 256 experts with 8 routed per token, so only about 3B of the 35B parameters are active for each token
- **Hybrid Attention** - Alternates linear-attention (Gated DeltaNet) layers with full attention, one full-attention layer in every four
- **Reasoning Model** - Opens each answer with a thinking block; the serving setup returns it in a separate reasoning field
- **Tool Calling** - Emits function calls that the server parses into standard OpenAI-style tool calls
- **Native 262K Context** - Handles 262,144 tokens natively, extensible to roughly 1M tokens with YaRN RoPE scaling
- **Image Input** - The checkpoint keeps the vision encoder from its Qwen3.5 base and accepts images alongside text

#### Use Cases

- Coding agents and coding CLIs pointed at an OpenAI-compatible endpoint
- Repository-level software engineering tasks and bug fixing
- Multi-step tool use and MCP-based agent workflows
- Long-context code understanding across large codebases
- Research and web-browsing agents

#### Benchmark Performance

The publisher reports the following results, averaged over five independent runs.

**Coding:**
- SWE-bench Verified: 79
- SWE-bench Multilingual: 71.4
- SWE-bench Pro: 59.6
- Terminal-Bench 2.1 (Terminus-2): 67.8
- Terminal-Bench 2.1 (Claude Code): 68.5
- NL2Repo: 46.2
- SWE Atlas QnA: 39.8
- DeepSWE: 22

**Reasoning:**
- GPQA Diamond: 89.2
- HLE (no tools): 25.6
- HLE (with tools): 33.4

**Agentic:**
- ClawEval: 72.5
- MCP-Atlas: 70.2
- WideSearch: 67.8
- BrowseComp: 67.6
- Toolathlon-Verified: 48.7

#### Recommended Sampling

The publisher recommends temperature 0.6, top_p 0.95 and top_k 20 for general tasks, and temperature 1.0 to reproduce the reported benchmarks.

Ornith 1.5 35B A3B gives teams a small-footprint agentic coding model: about 3B active parameters per token, a 262K native context window, and built-in reasoning and tool calling for coding agents. Model weights and the publisher's quantized builds are available on [HuggingFace](https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B).


Agentic coding MoE fine-tuned from Qwen3.5 with 3B active parameters and 262K context

Q5_K_M

Q4_K_M

Ornith 1.5 35B A3B


### Ornith 1.5 9B: Self-Improving Agentic Coding Model

Ornith-1.5-9B is the smallest member of the Ornith-1.5 family from Ornith AI, a 9B dense reasoning model built for agentic coding and tool use. The Ornith line started from Qwen3.5 and Gemma 4 base models and added its own continued pretraining, mid-training and post-training. This 9B checkpoint uses the Qwen3.5 architecture.

Ornith-1.5 is trained through end-to-end self-improvement. Rather than relying on a fixed set of human-curated tasks and hand-built agent harnesses, the training loop generates its own new tasks, builds the scaffolds used to attempt them, discovers strategies that solve them, and improves the model through reinforcement learning. Ornith-1.0 optimized scaffolds and rollouts; Ornith-1.5 extends the loop to task generation as well.

#### Key Features

-   **Reasoning model** - each answer opens with a thinking block, which the server returns separately from the final answer; the chat template also accepts an `enable_thinking` switch to skip it
-   **Tool calling** - emits structured function calls that the server turns into standard OpenAI-style tool calls, so it plugs into existing agent frameworks
-   **Built for coding agents** - tuned for terminal-based coding assistants and works with any OpenAI-compatible client or coding CLI
-   **Long context** - a 262,144-token context window, enough for large codebases and long agent sessions
-   **Lightweight** - the smallest model in the Ornith-1.5 family, a dense model designed for efficient deployment

#### Architecture

Ornith-1.5-9B keeps the Qwen3.5 hybrid design: 32 layers in which every fourth layer is full attention and the other three are linear-attention (gated delta) layers. The linear layers hold a fixed-size state instead of a growing key-value cache, so memory grows slowly as context gets longer. A vision encoder feeds image tokens into the same text backbone. Weights are published in BF16.

#### Benchmarks

Ornith reports every result as an average over five runs, and compares the 9B against its predecessor Ornith-1.0-9B and the Qwen3.5-9B base. On coding, it scores 70.6 on SWE-bench Verified (Qwen3.5-9B: 53.2), 47.5 on SWE-bench Pro (31.3), 54.4 on SWE-bench Multilingual (39.7) and 32.4 on NL2Repo (16.2). On Terminal-Bench 2.1 it reaches 46.2 with the Terminus-2 harness and 47 with Claude Code (Qwen3.5-9B: 21.3 and 18.9).

Agentic results follow the same pattern: 54.2 on MCP-Atlas, 41.2 on Toolathlon-Verified, 56.4 on BrowseComp, 59.5 on WideSearch and 66.5 on ClawEval. On reasoning it scores 86.4 on GPQA Diamond and 20.2 on Humanity's Last Exam without tools, rising to 30.5 with tools.

#### Use Cases

-   Backend model for terminal coding agents and coding CLIs
-   Software engineering agents that read a repository, edit code, run tests and iterate
-   Tool-calling and MCP-based agents for workflow automation
-   Research and browsing agents that chain many search and tool steps
-   Reasoning-heavy question answering where a visible thinking trace is useful

Full model details are available on the [Ornith-1.5-9B model card](https://huggingface.co/ornith-ai/Ornith-1.5-9B).


Lightweight 9B reasoning model from Ornith, built on Qwen3.5 and trained by self-improving reinforcement learning for agentic coding and tool use

Ornith 1.5 9B


### VoxCPM2: Tokenizer-Free Multilingual Text-to-Speech

VoxCPM2 is OpenBMB's tokenizer-free, diffusion autoregressive text-to-speech model. It speaks 30 languages, outputs 48 kHz audio, and was trained on over 2 million hours of multilingual speech data. No language tag is needed: you pass text in any supported language directly, and the model infers appropriate prosody and expressiveness from the content.

Beyond plain synthesis, VoxCPM2 can design a new voice from a written description, clone a voice from a short reference clip, and reproduce a speaker in fine detail when given the clip's transcript as well.

#### Key Features

-   **30 languages** - Arabic, Burmese, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, and Vietnamese, plus Chinese dialects including Sichuanese, Cantonese, Wu, and Northeastern Mandarin
-   **Voice design** - generate a novel voice from a natural-language description alone (gender, age, tone, emotion, pace) by putting the description in parentheses at the start of the text, with no reference audio required
-   **Controllable cloning** - clone any voice from a short clip, with optional style guidance to steer emotion, pace, and expression while preserving timbre
-   **Ultimate cloning** - provide reference audio together with its transcript for audio-continuation cloning that carries over the speaker's vocal nuances
-   **48 kHz output** - accepts 16 kHz reference audio and produces 48 kHz speech through the built-in super-resolution of its AudioVAE V2, with no external upsampler
-   **Streaming** - supports real-time streaming generation

#### Architecture

VoxCPM2 is built on the MiniCPM-4 backbone and works without a discrete audio tokenizer. Its pipeline runs a local encoder, a base language model, a residual language model, and a local diffusion transformer (LocEnc, TSLM, RALM, LocDiT) autoregressively over continuous speech representations, and AudioVAE V2 decodes the result into the waveform. The language model runs at a token rate of 6.25Hz.

#### Serving on Vast

This template serves VoxCPM2 with vLLM-Omni through the OpenAI-compatible audio speech endpoint, returning 48 kHz mono audio as WAV or as streaming PCM. Send plain text for zero-shot speech, start the text with a description in parentheses to design a voice, or include reference audio to clone a speaker. The bundled Model UI offers standard synthesis and reference-audio cloning.

#### Use Cases

-   Multilingual narration, dubbing, and localization across 30 languages
-   Character and persona voices created from a text description for games, audio drama, and assistants
-   Voice cloning for consistent narrators and branded voices from short recordings
-   Audiobook, podcast, and accessibility narration at 48 kHz
-   Fine-tuning a custom voice, which the model supports with full fine-tuning or LoRA from as little as 5 to 10 minutes of audio

Model weights are available on HuggingFace at [openbmb/VoxCPM2](https://huggingface.co/openbmb/VoxCPM2).


OpenBMB's tokenizer-free 48 kHz text-to-speech model with 30 languages, voice design from a text description, and voice cloning

VoxCPM2


### Voxtral Mini 4B Realtime: Streaming Speech Transcription

Voxtral Mini 4B Realtime 2602 is a multilingual, realtime speech-transcription model from Mistral AI that transcribes with a delay of under 500 milliseconds. It transcribes audio while it is still being spoken, which suits voice assistants, live subtitles, and meeting transcription.

On Vast, the model is served by vLLM's realtime API: a client opens a WebSocket session on `/v1/realtime`, streams audio in small chunks, and receives the transcript back as text deltas while the audio is still arriving. Recorded files can also be sent to the OpenAI-compatible `/v1/audio/transcriptions` endpoint.

#### Key Features

**Natively Streaming**
The model is built on a streaming architecture with a custom causal audio encoder, so it does not wait for a complete recording before it starts transcribing. Both the audio encoder and the language model use sliding-window attention, which lets a session run for long stretches of audio.

**Configurable Delay**
The transcription delay can be tuned to trade latency against accuracy, from 80 milliseconds up to 2.4 seconds. Mistral recommends 480 milliseconds as the balance between latency and accuracy.

**Multilingual**
The model supports 13 languages: Arabic, German, English, Spanish, French, Hindi, Italian, Dutch, Portuguese, Chinese, Japanese, Korean, and Russian.

**Compact**
At 4 billion parameters, the model is compact enough for realtime streaming.

#### Use Cases

- Private meeting transcription
- Live subtitles and captions
- Realtime voice assistants with speech understanding
- Transcribing live streams and calls as they happen

#### Architecture

Voxtral Mini 4B Realtime pairs a language model of about 3.4 billion parameters with an audio encoder of about 970 million parameters. The audio encoder was trained from scratch with causal attention, which is what makes streaming possible. One text token covers 80 milliseconds of audio.

#### Benchmarks

On the multilingual FLEURS benchmark, Mistral reports an average word error rate of 8.72 percent at a 480 millisecond delay, against 5.90 percent for its offline Voxtral Mini Transcribe 2.0. On long-form English, the gap is narrow: 3.17 percent on TEDLIUM against 2.86 percent offline, and 5.05 percent on Meanwhile against 4.08 percent. Full results are on the [model card](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602).


Multilingual streaming speech-to-text model from Mistral AI that transcribes live audio in 13 languages with under 500 ms of delay

Voxtral Mini 4B Realtime


### MiniMax M3: Native Multimodal MoE Model with 1M-Token Context

MiniMax M3 is a native multimodal model from MiniMax with a 1M-token context window. It has about 428B total parameters with about 23B activated per token, and it takes text, image, and video input. MiniMax built it for long-horizon coding and agentic work, and reports frontier-level results across long-horizon agentic benchmarks in both coding and cowork tasks.

#### Key Features

- **Native Multimodality** - Trained on mixed text, image, and video data from the very first step, for deeper fusion across modalities than an adapter added later
- **1M-Token Context** - Long documents, large repositories, and long agent trajectories fit in a single context
- **MiniMax Sparse Attention (MSA)** - A sparse attention operator built for million-token contexts; MiniMax reports 9x faster prefill and 15x faster decode than MiniMax M2 at 1M context, with per-token compute cut to 1/20
- **Sparse Mixture-of-Experts** - Only about 23B of the 428B parameters run on any given token
- **Three Reasoning Modes** - Thinking can be enabled, disabled for lower latency and higher throughput, or left adaptive so the model decides when extra reasoning helps
- **Coding and Cowork** - Strong on long-horizon agentic benchmarks for both software engineering and general cowork tasks

#### Use Cases

- Long-horizon agentic coding and repository-level software engineering
- Multi-step agent workflows with tool calling
- Image and video understanding alongside text
- Long-document and multi-document analysis across very long contexts
- Assistants that switch between fast direct answers and step-by-step reasoning per request

#### Architecture and Design

MiniMax M3 is a Mixture-of-Experts language model paired with a vision encoder for image and video input. The language model has 60 layers: the first 3 are dense and the remaining 57 route each token through 4 of 128 experts plus one shared expert. MiniMax Sparse Attention replaces standard grouped-query attention in the sparse layers, cutting the attention compute and memory cost of long contexts while preserving model quality. The model weights and model card are on [Hugging Face](https://huggingface.co/MiniMaxAI/MiniMax-M3).

#### Inference

The reasoning mode is chosen per request as enabled, adaptive, or disabled, with adaptive as the default. MiniMax recommends temperature 1.0 and top-p 0.95.

Deploy MiniMax M3 on Vast.ai for multimodal reasoning, agentic coding, and long-context work on flexible GPU infrastructure.


Natively multimodal 428B MoE with 23B active parameters, sparse attention, and a 1M-token context for coding and agentic work

NVFP4 (NVIDIA)

Q8_0 (bartowski)

Q5_K_M (bartowski)

Q4_K_M (bartowski)

MiniMax M3


### BiRefNet: High-Resolution Background Removal

BiRefNet (Bilateral Reference Network) is the official model from the paper "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" (CAAI AIR 2024). It separates a foreground subject from its background with pixel-accurate precision, producing a clean alpha mask and a transparent-background cutout from a single input image. It is one of the most widely used open models for background removal.

#### Overview

Dichotomous image segmentation is the task of splitting an image into exactly two regions: the object of interest and everything else. BiRefNet approaches this with a bilateral reference architecture that combines an "inward" reference, which keeps the original high-resolution image detail available to the decoder, with an "outward" reference based on gradient information that guides the network's attention toward fine structures. The result is a model that preserves details that defeat most segmentation models, such as hair strands, fur, mesh, wires, and semi-transparent edges.

#### Key Features

-   **Precise segmentation masks** - Produces high-quality binary masks and soft alpha mattes, handling fine structures such as hair, netting, and thin branches
-   **High-resolution processing** - Standard inference runs at 1024x1024, with sibling checkpoints trained for resolutions up to 2048x2048
-   **State-of-the-art accuracy** - The paper reports leading performance on the three benchmark tasks it targets: dichotomous image segmentation (DIS5K), high-resolution salient object detection (HRSOD), and concealed object detection (COD)
-   **Lightweight and fast** - Compact model that loads in seconds and processes images in a fraction of a second on a modern GPU
-   **Model zoo** - The general-purpose checkpoint used here is one of a family that includes portrait, matting, and lite variants, all published under the same ZhengPeng7 namespace on Hugging Face

#### Use Cases

-   E-commerce product photography: cut out products for catalog images with clean edges
-   Portrait and profile picture background removal
-   Graphic design and marketing: extract subjects for compositing into new scenes
-   Photo editing pipelines: generate alpha mattes for downstream compositing tools
-   Batch background removal via the ComfyUI API for automated media workflows
-   Preprocessing for image generation: isolate subjects before inpainting or restyling

#### Architecture

BiRefNet pairs a Swin Transformer backbone with a decoder built around the bilateral reference mechanism. During decoding, patches of the source image at full resolution are fed back into each stage (the inward reference) so that reconstruction never has to rely on downsampled features alone, while gradient-aware supervision (the outward reference) pushes the model to attend to regions dense in fine detail. Auxiliary gradient supervision further sharpens its handling of edges and thin structures.

#### Community and Ecosystem

BiRefNet is among the most downloaded image segmentation models on Hugging Face, and its checkpoints power background removal features in many downstream tools and interfaces. The author maintains an active model zoo with task-specific weights for standard dichotomous segmentation, salient object detection, and concealed object detection, alongside the general-purpose checkpoint served here from [ZhengPeng7/BiRefNet](https://huggingface.co/ZhengPeng7/BiRefNet). This template loads it through ComfyUI's built-in background-removal nodes, using the ComfyUI-ready copy published at [Comfy-Org/BiRefNet](https://huggingface.co/Comfy-Org/BiRefNet), and saves each result as a transparent PNG.


State-of-the-art background removal and dichotomous image segmentation model producing clean masks and transparent foregrounds

BiRefNet


### Chatterbox: Open Text-to-Speech with Voice Cloning

Chatterbox is Resemble AI's production-grade open source text-to-speech family, published at [ResembleAI/chatterbox](https://huggingface.co/ResembleAI/chatterbox). From a short reference clip it clones a voice zero-shot, and it includes an emotion exaggeration control that dials expressiveness up or down independently of the text. The multilingual checkpoints extend the same zero-shot cloning across 23 languages including Arabic, Chinese, French, German, Hindi, Japanese, Korean, Portuguese, Russian, Spanish, and Swahili.

#### Overview

The core model pairs a 0.5B-parameter Llama-style autoregressive backbone with the S3Gen token-to-waveform generator and a compact voice encoder that conditions generation on the reference speaker. Training on roughly half a million hours of cleaned speech, combined with alignment-informed inference, keeps long generations stable with little of the drift, repetition, or skipped words that plague autoregressive TTS. Two creative controls set Chatterbox apart: an exaggeration slider that pushes delivery from flat narration to dramatic performance, and a classifier-free guidance weight that trades pacing against adherence, so a fast-talking reference can be slowed into deliberate speech.

#### Key Features

- **Zero-shot voice cloning** - A few seconds of reference audio is enough to synthesize new speech in that voice, with no fine-tuning
- **Emotion exaggeration control** - A single parameter scales expressive intensity
- **Multilingual synthesis** - The multilingual checkpoint speaks 23 languages and supports cross-language cloning, where a voice recorded in one language speaks another
- **Turbo variant with paralinguistic tags** - The faster Turbo checkpoint renders inline tags such as laughs, sighs, gasps, and throat-clearing for natural conversational delivery
- **Voice conversion** - A dedicated conversion pipeline re-voices existing recordings into a target speaker
- **Stable long-form generation** - Alignment-informed inference keeps output on-script for agent, narration, and video workloads

#### Use Cases

- Voice agents and conversational AI that need low-latency, consistent speech
- Narration for videos, games, memes, and audiobooks in a cloned or house voice
- Localization: one recorded voice speaking all 23 supported languages
- Multi-speaker dialog synthesis for drama, podcasts, and game NPCs
- Re-voicing existing audio through voice conversion
- Accessibility tools and screen-reader-quality reading voices

#### On This Template

The template serves Chatterbox through ComfyUI with the FL ChatterBox node pack preinstalled and every checkpoint pre-downloaded: standard English TTS, the 23-language multilingual model, the Turbo model with paralinguistic tag support, and the voice conversion model. Three ready-made workflows are preloaded, one each for English, multilingual, and Turbo synthesis. Each runs immediately with the built-in reference voice; connect any short audio clip to the optional audio prompt input to clone a voice instead. Dialog and voice conversion nodes from the same pack are available to wire in without downloading anything else. As in Resemble AI's own release, every generated clip carries Resemble AI's imperceptible Perth watermark, which lets the audio be identified as AI-generated.


Resemble AI's production-grade open TTS with zero-shot voice cloning, emotion exaggeration control, and multilingual synthesis across 23 languages

Chatterbox


### DeepSeek V4.1 Flash

DeepSeek V4.1 Flash is a multimodal Mixture-of-Experts (MoE) language model with 552B backbone parameters and support for contexts of up to one million tokens. It natively processes images and text and generates text. The model activates only 8B parameters per token during prefill and 16B during decode, which substantially improves cost efficiency for input-heavy agentic workloads.

### Architecture

DeepSeek V4.1 Flash adopts a Causal Encoder-Decoder design: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. The decoder's global KV cache is projected from the final encoder hidden states rather than computed at every decoder layer, which is where the split between prefill and decode activation comes from.

Attention uses Compressed Sparse Attention 2 (CSA2), which assigns each layer a static mode that shares the main KV and indexer keys across layers and reuses sparse-attention indices. Combined with FP4 KV caching, this brings the global KV cache to about 890 bytes per token, roughly a quarter of DeepSeek V4 Flash. Further components include Engram conditional memory (196B parameters accessed sparsely through token-based lookup), a revised single-pass residual mixing scheme, and a DSpark speculative decoding module. Each MoE layer has one shared expert and 384 routed experts, with 6 routed experts active per token.

A vision encoder trained from scratch and a two-layer projector turn images into visual embeddings, which the model processes jointly with text from the start of pre-training. The model was trained from scratch on a 45T-token multimodal corpus, with context extended to one million tokens.

### Key Features

- Multimodal MoE model with 552B backbone parameters, 8B active per token in prefill and 16B in decode
- Context window of up to one million tokens with a very small KV cache footprint
- Native image understanding alongside text
- Continuously controllable reasoning effort from 1 to 100, trading inference cost for accuracy
- Native tool calling and reasoning parsing for agentic applications
- Open-source under the MIT License

### Use Cases

- Coding agents and software engineering tasks
- Terminal and tool-using agents that run long multi-step tasks
- Long-document analysis and retrieval over very large inputs
- Visual question answering, chart and document understanding
- Mathematical problem-solving and step-by-step reasoning

### Benchmarks

At maximum reasoning effort, DeepSeek reports 90.6 on Terminal-Bench 2.1, 74.2 resolved on DeepSWE v1.1, 64.0 on NL2Repo-Bench, 88.1 on CyberGym, 63.9 on HLE with tools, 54.8 on AutomationBench and 31.8 on Agent's Last Exam. On reasoning benchmarks it scores 90.9 on GPQA Diamond, a Codeforces rating of 3471, and 65.6 on MathArena Apex.

These agentic results are well ahead of DeepSeek V4 Flash, which scored 82.7 on Terminal-Bench 2.1 and 54.4 on DeepSWE v1.1. With image input, the model scores 78.9 on Chartography and 89.6 on BabyVision, both with tools.

For full model details and the complete evaluation tables, see the model card on [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash).


Multimodal Mixture-of-Experts agentic model with 1M context

DeepSeek V4.1 Flash


### GLM 5.3: Agentic Coding and Long-Horizon Model

GLM 5.3 is a Mixture-of-Experts model from Z.ai. It uses the same base model as GLM 5.2, and every gain comes from post-training. Compared with GLM 5.2 it is much stronger at complex coding and long-horizon agentic tasks, with native English and Chinese support.

#### Key Features

- **Stronger Coding** - Z.ai reports a 50% improvement over GLM 5.2 on its in-house Z.ai Code Bench
- **Terminal and Agent Work** - 88.2 on Terminal Bench 2.1 (up from 81.0) and 28.3 on Terminal Bench 3.0 (up from 4.6)
- **Software Engineering** - 66.9 on DeepSWE, up from 46.2 for GLM 5.2
- **Security Research** - 84.5 on CyberGym for vulnerability discovery
- **Tool Use** - 62.5 on Humanity's Last Exam with tools and 73.0 on Toolathlon Verified
- **Adjustable Thinking** - The `reasoning_effort` setting accepts `low`, `high`, or `max` (the default)

#### Use Cases

- Repository-level software engineering and multi-file code changes
- Long-running agentic workflows in terminal and coding harnesses
- Vulnerability discovery and security analysis
- Multi-step tool calling and automation
- Long-context document and codebase analysis

#### Architecture and Design

GLM 5.3 keeps the GLM 5.2 architecture: a sparse-attention Mixture-of-Experts design in which a lightweight indexer picks the tokens each query attends to, keeping long-context inference affordable across its 1M-token window. The checkpoint ships in FP8 and includes a multi-token prediction layer.

#### Training and Inference

All improvements over GLM 5.2 come from post-training. Z.ai evaluates with a temperature of 1.0 and top-p of 0.95. In the chat template, `clear_thinking` defaults to false; for chat use, pass `clear_thinking=true`. Leave `reasoning_effort` at its default `max` to reproduce published benchmark results.

Deploy GLM 5.3 on Vast.ai for agentic coding and long-context work on flexible GPU infrastructure.


753B MoE model tuned for complex coding and long-horizon agentic work, with 1M-token context

UD-Q4_K_XL (Unsloth)

GLM 5.3


### Kokoro 82M: Small, Fast, Open Text-to-Speech

Kokoro is an open-weight text-to-speech model with just 82 million parameters, published at [hexgrad/Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M). Despite its size, it delivers speech quality comparable to models many times larger: entered into the community TTS arenas shortly after release, it climbed to the top of the rankings while remaining fast and cheap enough to run interactively on modest hardware. The v1.0 release ships 54 voices across 8 languages.

#### Overview

Kokoro's design philosophy is efficiency. It builds on the StyleTTS 2 architecture with an ISTFTNet vocoder, using a decoder-only formulation with no diffusion process, so synthesis is a single fast forward pass rather than an iterative sampling loop. The entire model was trained for roughly a thousand A100 GPU-hours on a few hundred hours of permissively licensed and synthetic audio with IPA phoneme labels. Text is converted to phonemes by the companion misaki G2P library, and the model generates 24 kHz audio conditioned on a chosen voice embedding.

#### Key Features

- **Tiny footprint** - 82M parameters; loads in seconds and synthesizes far faster than real time on any modern GPU
- **54 voices, 8 languages** - American and British English, Spanish, French, Hindi, Italian, Brazilian Portuguese, Japanese, and Mandarin Chinese, each with multiple named voices
- **Voice blending** - Voice embeddings can be interpolated, mixing two voices at an adjustable ratio to create new ones
- **No diffusion** - Decoder-only StyleTTS 2 + ISTFTNet design produces audio in one pass, keeping latency low and throughput high
- **Truly open** - Open weights trained only on permissive, non-copyrighted, and synthetic audio, with commercial use encouraged

#### Architecture

Kokoro pairs the StyleTTS 2 formulation with an ISTFTNet-based vocoder that reconstructs the waveform through inverse short-time Fourier transform layers rather than a heavy neural upsampling stack. A style vector selected from the voice pack conditions prosody and timbre, while the phoneme sequence from misaki drives content. Because there is no diffusion sampler and no large acoustic encoder at inference time, generation cost scales gently with text length, which is what lets an 82M-parameter model serve production traffic efficiently.

#### Use Cases

- Voiceover and narration: turn scripts, articles, or documentation into natural-sounding audio
- Application and game voices: low-latency speech for assistants, NPCs, and accessibility features
- Audiobook and podcast production: long-form synthesis where per-character cost matters
- Multilingual content: one model covering eight languages with consistent quality
- Custom voice design: blend existing voices to produce a distinct brand voice without training

#### On This Template

The template runs Kokoro inside ComfyUI through a dedicated TTS node pack. A ready-made workflow is preloaded: type or paste text, pick from the built-in voice list, adjust the speaking speed, and queue - the synthesized speech is saved as a FLAC file you can play directly from the output panel. The node also exposes voice blending, mixing any two voices at a chosen ratio. The template ships 41 voices across 7 languages: American and British English, Spanish, French, Hindi, Italian, and Brazilian Portuguese. The Japanese and Mandarin voices are not included, because their text-processing libraries are not installed. The model weights and voice packs are downloaded during provisioning, so the first generation starts immediately. The same workflow is also available through the included API wrapper for programmatic and serverless use.


Compact 82M-parameter open-weight text-to-speech model with 50+ voices across 8 languages, delivering natural speech quality that rivals far larger models at a fraction of the cost

Kokoro 82M


### LFM2.5 VL 3B: Compact Vision-Language Model from Liquid AI

LFM2.5 VL 3B is a 3.1B-parameter vision-language model from Liquid AI. It reads images and text and answers in text. It pairs the LFM2.5-2.6B hybrid language model with a SigLIP2 NaFlex vision encoder, and builds on LFM2-VL-3B with further mid- and post-training.

#### Key Features

-   **Grounding and object detection** - Finds objects from natural-language queries and returns boxes as normalized coordinates
-   **Full-page OCR with layout** - Parses documents into labelled regions such as text, titles, tables, charts and equations
-   **Native resolution** - Large images are split into 512x512 patches plus a whole-image thumbnail instead of being squashed
-   **Tool calling** - Writes Pythonic function calls that the server returns as standard tool calls
-   **Multilingual** - Supports 16 languages, including English, Chinese, Japanese, Korean, Arabic, Hindi and the main European languages
-   **32K context** - Room for multi-image prompts and long documents

#### Use Cases

-   Batch OCR of scanned documents into searchable, structured text
-   Screen and UI understanding
-   Near-realtime object detection and grounding
-   Translating photographed menus and signs
-   Visual question answering and image description

#### Architecture

The language backbone is LFM2.5-2.6B, a hybrid model where most layers are short gated convolutions and a minority are grouped-query attention. That keeps memory for long context small. The vision side is a 400M SigLIP2 NaFlex encoder. The vocabulary is 128,000 tokens.

The model answers directly rather than reasoning step by step, so time to first token is short. Liquid recommends it for single-turn, high-throughput, low-latency tasks, and not for long-context, reasoning-heavy work.

#### Performance

Liquid reports large gains over LFM2-VL-3B in screen understanding, grounding, multi-image input and tool use: 80.7 on ScreenSpot-v2, 87.9 on RefCOCO, 61.5 on BLINK, 58.3 on MuirBench and 59.5 on ToolSandBox. On general vision benchmarks it scores 73.1 on RealWorldQA, 81.3 on ChartQA, 68.5 on MathVista and 88.7 on POPE.

Recommended text sampling is temperature 0.2 and top_k 50.

Learn more on the [model card](https://huggingface.co/LiquidAI/LFM2.5-VL-3B).


Compact 3B vision-language model from Liquid AI with 32K context, OCR, grounding and tool calling

LFM2.5 VL 3B


### RIFE: Real-Time Video Frame Interpolation

RIFE (Real-Time Intermediate Flow Estimation) is the frame interpolation model from the ECCV 2022 paper "Real-Time Intermediate Flow Estimation for Video Frame Interpolation" by Huang et al. Given two consecutive video frames, it synthesizes new frames in between, so a clip can be converted to a higher frame rate for smoother motion or stretched into fluid slow motion. It is the most widely used open frame interpolation model, powering interpolation features in many video enhancement tools and generative video pipelines.

#### Overview

Classical frame interpolation methods first estimate bidirectional optical flow between the two input frames and then warp and blend them, which is slow and produces artifacts at motion boundaries. RIFE instead trains a dedicated network, IFNet, to estimate the intermediate flows directly - the flows from the in-between moment to each of the two input frames - in a single coarse-to-fine pass. A privileged distillation scheme, where a teacher that can peek at the ground-truth middle frame supervises the student, stabilizes training and sharpens results. The result is a model that runs in real time on ordinary GPUs while remaining competitive in quality with much heavier interpolation networks.

This template serves RIFE 4.9 from the practical 4.x line of RIFE checkpoints, which the author maintains as an actively improved successor to the original paper release, tuned for robustness on real-world footage and animation rather than only on benchmarks.

#### Key Features

-   **Arbitrary-timestep interpolation** - The 4.x models synthesize a frame at any point between two inputs, enabling 2x, 4x, 8x, or non-integer frame rate conversion from a single model
-   **Real-time performance** - The model is compact and interpolates HD frames in a fraction of a second, so full clips process quickly
-   **Strong motion handling** - End-to-end intermediate flow estimation keeps motion boundaries clean where warp-and-blend approaches produce ghosting
-   **Works on real footage and animation** - The practical checkpoint line is trained for robustness across live action, rendered content, and anime

#### Use Cases

-   Frame rate upscaling: convert 24 or 30 fps footage to 60 fps or higher for smooth playback
-   Slow motion: multiply frames while keeping the original frame rate to stretch a clip in time
-   Smoothing AI-generated video: interpolate the low-fps output of text-to-video and image-to-video models into fluid clips
-   Restoring choppy or low-frame-rate sources such as animation, screen recordings, and archival footage
-   Batch video processing through the ComfyUI API for automated pipelines

#### Architecture

RIFE is built around IFNet, a lightweight fully convolutional network that estimates intermediate optical flow coarse-to-fine across three resolution scales, refining the flow field and a fusion mask at each stage. The two input frames are backward-warped by the estimated flows and blended by the mask, and a small refinement module cleans up the residual details. Because the flow is predicted directly for the target timestep rather than derived from bidirectional flow, the whole pipeline is a single feed-forward pass, which is what makes real-time operation possible.

#### On This Template

The template runs RIFE inside ComfyUI using the ComfyUI-Frame-Interpolation node pack, with the RIFE 4.9 checkpoint (rife49.pth) preloaded. A ready-made workflow is preloaded along with a sample clip: the video is split into frames, the RIFE VFI node doubles them, and the result is reassembled and saved at twice the original frame rate. Set the multiplier on the RIFE VFI node for 4x or higher interpolation, and set the frame rate on the Create Video node to match: source frame rate times multiplier for smooth motion, or the original frame rate for slow motion. Keep rife49.pth selected in the RIFE VFI node: the node pack's download of rife426.pth currently fails to load (a known issue in the node pack). Upload your own video in the Load Video node to process your footage, and use the bundled API wrapper payload for programmatic interpolation jobs.


Real-time video frame interpolation model that doubles or multiplies frame rate for smooth motion and slow-motion effects

RIFE


### SAM 2: Segment Anything in Images and Videos

Segment Anything Model 2 (SAM 2) is Meta FAIR's foundation model for promptable visual segmentation in images and videos, introduced in the paper "SAM 2: Segment Anything in Images and Videos". Given a prompt — a click, a box, or a mask — it produces a precise segmentation mask for the indicated object, and it can also segment everything in an image automatically with no prompt at all. This template runs the SAM 2.1 Hiera Large checkpoint, the largest and most accurate of the improved SAM 2.1 weights published at [facebook/sam2.1-hiera-large](https://huggingface.co/facebook/sam2.1-hiera-large).

#### Overview

SAM 2 extends the original Segment Anything Model from images to video by treating an image as a one-frame video. The design is a simple transformer with streaming memory: each frame is encoded once, and a memory attention module conditions the current frame on the ones before it, so an object selected on one frame can be tracked and re-segmented across an entire clip in real time. A model-in-the-loop data engine, where annotators and the model refine each other, was used to build the SA-V dataset, the largest video segmentation dataset released to date, and SAM 2 trained on it delivers strong performance across a wide range of tasks and visual domains.

#### Key Features

- **Promptable segmentation** - Point clicks (positive and negative), bounding boxes, and input masks all steer the prediction; a few clicks are usually enough to isolate a subject cleanly
- **Automatic mask generation** - With no prompt, the automatic mask generator sweeps a grid of points across the image and returns a mask for every distinct object it finds
- **Zero-shot generality** - As a segmentation foundation model, it handles objects and scenes it was never explicitly trained on, without fine-tuning
- **Video object tracking** - The streaming-memory architecture propagates masks across frames, tracking one or many objects through a video from prompts on a single frame
- **A family of checkpoints** - SAM 2.1 ships tiny, small, base-plus, and large variants; Hiera Large, served here, is the most accurate of the four while remaining fast enough for interactive use

#### Architecture

SAM 2 pairs a hierarchical Hiera vision transformer encoder with a lightweight prompt encoder and mask decoder derived from the original SAM design. For video, a memory encoder compresses previous frames' predictions into a memory bank, and memory attention lets the current frame attend to it, which is what allows masks to follow an object through occlusion and motion. The SAM 2.1 checkpoints are a retrained suite with improved accuracy over the original SAM 2 release; the Hiera Large variant used here has 224.4 million parameters, so it loads in seconds and segments an image in a fraction of a second on a modern GPU.

#### Use Cases

- Interactive object cutouts: click a subject to get a clean mask for compositing or editing
- Automatic scene decomposition: segment every object in an image for analysis or labeling pipelines
- Dataset annotation: bootstrap segmentation labels far faster than manual polygon tracing
- Masking for generative editing: feed SAM 2 masks into inpainting and regional-prompt workflows
- Video rotoscoping and object tracking: select an object once and propagate its mask through a clip
- Robotics and vision research: a strong off-the-shelf segmentation signal for downstream systems

#### On This Template

The template runs SAM 2.1 Hiera Large inside ComfyUI via dedicated SAM 2 nodes. A ready-made workflow is preloaded: load any image and the automatic mask generator segments every object it finds, saving a color-coded segmentation overlay and a black-and-white mask of everything it found. The workflow is also available through the built-in API wrapper, which returns both images. The same node pack also provides point- and box-prompted segmentation and video segmentation nodes, so the preloaded checkpoint can be wired into selective masking or tracking workflows without downloading anything else.


Meta's Segment Anything Model 2.1 for promptable object segmentation in images and video, producing precise masks from clicks, boxes, or fully automatic mask generation

SAM 2.1 (Hiera Large)


# Qwen3 Embedding

Qwen3 Embedding is the Qwen family's text-embedding model series, purpose-built for embedding and ranking tasks such as text retrieval, code retrieval, semantic search, text classification, clustering, and bitext mining. Built on the dense Qwen3 foundation models, the series inherits their multilingual capability, long-text understanding, and reasoning skill, and ships in three sizes — 0.6B with 1024-dimension embeddings, 4B with 2560 dimensions, and 8B with 4096 dimensions — all with a 32K-token context window.

On Vast, the flagship template serves [Qwen3-Embedding-8B](https://huggingface.co/Qwen/Qwen3-Embedding-8B) through vLLM's pooling runner, exposing an OpenAI-compatible embeddings endpoint at `/v1/embeddings`. The 4B and 0.6B variants trade a few points of quality for lower cost and latency on smaller GPUs.

## Key Features

- **State-of-the-art retrieval quality.** The 8B embedding model ranked No.1 on the MTEB multilingual leaderboard as of June 5, 2025, with a mean task score of 70.58; the 4B model scores 69.45 and the 0.6B model 64.33 on the same benchmark.
- **Flexible embedding dimensions (MRL).** All sizes support Matryoshka Representation Learning: user-defined output dimensions from 32 up to each model's native width (4096 for 8B, 2560 for 4B, 1024 for 0.6B), so you can shrink vectors to fit your index without retraining.
- **Instruction aware.** Queries can carry a one-sentence task instruction (for example, "Given a web search query, retrieve relevant passages that answer the query"). Qwen's evaluations show instructions typically improve downstream task quality by 1 to 5 percent; they recommend writing instructions in English.
- **Multilingual and code-capable.** The series supports over 100 languages, including programming languages, with strong multilingual, cross-lingual, and code-retrieval performance.
- **Long inputs.** Up to 32K tokens per input, suitable for embedding whole documents rather than fragments.

## Use Cases

- Retrieval-augmented generation (RAG) pipelines and vector databases
- Semantic search across multilingual corpora
- Code search and cross-lingual retrieval
- Text classification, clustering, and deduplication
- Reranking pipelines, paired with the companion Qwen3-Reranker models

## Architecture

Each size is a dense Qwen3 causal-language-model backbone finetuned for embedding: the 0.6B model has 28 layers, and the 4B and 8B models have 36 layers. Embeddings are produced by pooling over the final hidden states, then normalized. Serving through vLLM's pooling runner converts the causal-LM checkpoint into an embedding model at load time.

## Benchmarks

On the MTEB multilingual benchmark (evaluated May 24, 2025), Qwen3-Embedding-8B leads with a mean task score of 70.58, ahead of gemini-embedding-exp-03-07 at 68.37, and well clear of text-embedding-3-large at 58.93 and Cohere-embed-multilingual-v3.0 at 61.12. Qwen3-Embedding-4B reaches 69.45 — also ahead of every compared non-Qwen model — and even the 0.6B model's 64.33 outscores text-embedding-3-large and Cohere's multilingual embedder. On MTEB English v2 the 8B model scores 75.22 mean task, with the 4B at 74.60 and the 0.6B at 70.70. Full results are on the [Qwen3-Embedding-8B model card](https://huggingface.co/Qwen/Qwen3-Embedding-8B).


Multilingual text-embedding family (0.6B-8B) serving OpenAI-compatible /v1/embeddings via vLLM

0.6B

Qwen3 Embedding


### Whisper: Robust Speech Recognition and Translation

Whisper is a family of automatic speech recognition and speech translation models from OpenAI, introduced in the paper Robust Speech Recognition via Large-Scale Weak Supervision. The flagship large-v3 checkpoint was trained on 1 million hours of weakly labeled audio and 4 million hours of pseudo-labeled audio, and delivers a 10 to 20 percent error reduction over large-v2.

On Vast, Whisper is served through the OpenAI-compatible audio transcription endpoint, so any client that already speaks the OpenAI transcription API can point at the instance without code changes.

#### Key Features

**Multilingual Transcription**
Whisper transcribes speech in 99 languages, with automatic language detection. Robustness to accents, background noise, and technical language comes from the scale and diversity of its training data, and it generalizes zero-shot to many domains without fine-tuning.

**Speech Translation**
Beyond same-language transcription, Whisper translates speech from any supported language directly into English text in a single pass.

**Timestamps**
The model produces sentence-level and word-level timestamps, making it a complete pipeline from raw audio to captions and searchable transcripts.

**Drop-in API**
Transcription requests use the same request and response shape as the OpenAI audio transcription API, including the choice of plain text, JSON, or subtitle-style output.

#### Use Cases

- Transcribing meetings, interviews, lectures, and podcasts
- Generating subtitles and captions for video content
- Translating foreign-language audio into English text
- Building voice-driven applications on a private transcription API
- Batch transcription of audio archives

#### Architecture

Whisper is a Transformer-based encoder-decoder model with 1.55 billion parameters in its large-v3 configuration. Audio is converted to a 128-bin Mel spectrogram, up from 80 bins in large-v2, and processed in 30-second windows, with long-form audio handled through sequential or chunked decoding. On the Open ASR Leaderboard, large-v3 reports a mean word error rate of 7.44 across evaluation datasets.


Whisper transcribes and translates speech in 99 languages, served through the OpenAI-compatible audio transcription endpoint

Whisper large-v3


### Depth Anything 3: Visual Geometry from Any Views

Depth Anything 3 (DA3) is the visual geometry model series from the ByteDance Seed team, released in November 2025 as the successor to Depth Anything V2. Where V2 predicts a disparity map from a single image, DA3 predicts depth directly and recovers spatially consistent geometry from arbitrary visual inputs — single images, multi-view sets, or video — with or without known camera poses. This template runs the [DA3MONO-LARGE](https://huggingface.co/depth-anything/DA3MONO-LARGE) checkpoint, the monocular-depth-optimized member of the family, using the safetensors conversion published at [Comfy-Org/Depth-Anything-3](https://huggingface.co/Comfy-Org/Depth-Anything-3).

#### Key Ideas

The DA3 paper is built on two findings. First, a single plain transformer — a vanilla DINO encoder with no architectural specialization — is a sufficient backbone for visual geometry. Second, a unified depth-ray representation lets one model family cover monocular depth, multi-view depth, camera pose estimation, and 3D reconstruction without complex multi-task learning. The models are trained on public academic datasets only, with a teacher-student pipeline in the tradition of the earlier Depth Anything releases.

#### Key Features

- Direct depth prediction rather than V2-style disparity, giving superior geometric accuracy
- Video-consistent depth: frames processed with cross-view attention stay temporally stable instead of flickering
- Sky segmentation output alongside the depth map
- Significantly outperforms Depth Anything 2 on monocular depth estimation, and the wider DA3 family surpasses VGGT on multi-view depth and camera pose benchmarks
- Compact single-file checkpoint at 0.35B parameters

#### The Model Family

The series spans Small, Base, Large, and Giant any-view geometry models plus specialized variants: DA3MONO-LARGE for monocular relative depth, DA3METRIC-LARGE for metric depth, and a nested Giant-Large teacher model. The Mono-Large checkpoint served here is the variant ComfyUI's official Depth Anything 3 workflows are built around, tuned for the highest-quality single-image and per-frame depth.

#### Use Cases

- Depth maps for 3D parallax, relighting, depth-of-field, and compositing effects
- Temporally consistent depth for video editing and VFX pipelines
- Depth conditioning for image and video generation workflows
- Robotics, drone, and navigation prototyping from a monocular camera signal
- Preprocessing for point clouds, novel view synthesis, and 3D reconstruction

#### On This Template

The template runs Depth Anything 3 inside ComfyUI using the native DA3 nodes, with the two official workflows preloaded. The image workflow takes any picture and produces a normalized depth map; the video workflow runs consistent depth over every frame of an uploaded clip and renders the result back to video. Outputs can be wired into other ComfyUI nodes or exported for use in external tools.


ByteDance's visual-geometry model for monocular and video-consistent depth estimation, recovering spatially consistent geometry from any visual input

Depth Anything 3 (Mono Large)


### MiMo V2.6 Distill Qwen 9B: Agentic Coding and Tool Use

MiMo-V2.6-Distill-Qwen-9B is an agentic model from Xiaomi MiMo, built by supervised fine-tuning of [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) on data generated by the larger MiMo-V2.6 family. It targets four areas in particular: software engineering, general-purpose agent work, visual coding, and cybersecurity. Xiaomi released it as an SFT checkpoint intended as a starting point for open research into agentic reinforcement learning, so it is as useful as an RL base as it is as a served assistant.

#### Key Features

-   **Agentic by design** - trained on multi-step tool-use and terminal trajectories rather than single-turn chat, so it plans, calls tools, and recovers from failed steps
-   **Explicit thinking mode** - the chat template exposes an `enable_thinking` switch, and reasoning content is returned separately from the final answer
-   **Multimodal input** - accepts images alongside text, which is what powers its visual coding ability (reading a screenshot or mockup and writing the code for it)
-   **Long context** - a 262,144-token context window, enough to hold large repositories, long agent traces, and extended tool transcripts in a single session
-   **Strong code and security focus** - the training mixture is weighted heavily toward code and cybersecurity tasks rather than general web text

#### Architecture

The model inherits the Qwen3.5 architecture: a hybrid attention stack that interleaves three linear-attention (gated delta) layers with one full-attention layer, repeating through all 32 layers. The linear layers carry a fixed-size recurrent state instead of a growing key-value cache, so memory at long context grows far more slowly than in a conventional transformer of similar depth. A vision encoder feeds image tokens into the same text backbone, giving one unified model for text and image input rather than a bolted-on adapter. Weights are published in BF16.

#### Benchmarks

Xiaomi reports results for this SFT checkpoint against its Qwen3.5-9B starting point in the MiMo-V2.6 technical report. The gains are largest where agentic training matters most. On SWE Pro it scores 44.6 against 32.0 for the base model, and on the internal MiMo Code evaluation it reaches 51.6 against 19.5. Cybersecurity shows the widest gap, 31.3 against 5.7 on the internal MiMo Cyber set. General agent benchmarks move similarly: AutomationBench v1.0.6 rises to 30.3 from 5.0, Terminal Bench 2.1 to 37.1 from 27.0, Toolathlon-Verified to 35.2 from 25.9, and JobBench to 18.3 from 2.6. On SWE Verified the two are close, 61.1 against 60.0, and on visual coding the model reaches 64.0 against 61.7.

#### Training Data

The supervised fine-tuning mixture totals 77.4 billion tokens, of which 27.2 billion are loss-bearing. It is split roughly evenly across four domains: code contributes the largest share at about 30 percent of tokens, followed by general agent data at about 28 percent, visual data at about 27 percent, and cybersecurity at about 14 percent. The visual and code slices carry the most loss-bearing tokens, which matches where the reported gains concentrate.

#### Use Cases

-   Autonomous software engineering agents that read a repository, edit files, run tests, and iterate
-   Terminal and shell agents that chain commands and handle their own error recovery
-   Visual coding, turning a screenshot, design mockup, or diagram into working code
-   Security research assistants for capture-the-flag style tasks and vulnerability triage
-   Tool-calling backends for workflow automation across office documents and internal APIs
-   A starting checkpoint for teams running their own agentic reinforcement learning experiments

Full model details are available on the [MiMo-V2.6-Distill-Qwen-9B model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B).


Agentic 9B vision-language model distilled from Qwen3.5-9B for coding, tool use, and cybersecurity

MiMo V2.6 Distill Qwen 9B


MinerU2.5-Pro is a 1.2B-parameter document parsing vision-language model from OpenDataLab, the team behind the MinerU PDF-to-Markdown toolchain. It converts page images of PDFs and scanned documents into structured output — Markdown text, HTML tables, and LaTeX formulas — and is the current flagship checkpoint of the MinerU 2.5 line ([opendatalab/MinerU2.5-Pro-2605-1.2B](https://huggingface.co/opendatalab/MinerU2.5-Pro-2605-1.2B)).

## Key Features

MinerU2.5 adopts a coarse-to-fine, two-stage parsing strategy: it first performs global layout analysis on a downsampled view of the page, then runs fine-grained content recognition on native-resolution crops of each detected block. This decoupled design lets a small 1.2B model handle high-resolution documents with low computational overhead while preserving non-body elements such as headers, footers, and page numbers.

The Pro checkpoint was produced purely through data engineering on the same 1.2B architecture: the training corpus was expanded from under 10 million to 65.5 million pages with difficulty-aware sampling, and annotations for complex tables and dense formulas were regenerated with cross-model consistency verification and an iterative judge-and-refine pipeline, followed by three-stage progressive training ending in GRPO format alignment.

Beyond core parsing, MinerU2.5-Pro adds practical capabilities for production document pipelines: image and chart analysis, truncated paragraph merging across blocks, in-table image recognition, and improved handling of rotated, borderless, and partially bordered tables. Mixed-language Chinese-English formulas and long, complex equations are parsed to LaTeX.

## Architecture

The model is built on the Qwen2-VL architecture (Qwen2VLForConditionalGeneration) with a vision encoder feeding a compact 1.2B language decoder, and ships in Safetensors format on the Transformers framework. It serves with vLLM out of the box; the publisher recommends vLLM for inference and reports concurrent throughput of about 2 frames per second on a single data-center GPU with the asynchronous engine.

Each parsed page is returned as a list of content blocks with a type (text, table, equation, or image), a normalized bounding box, an optional rotation angle, and the recognized content. The publisher's mineru-vl-utils Python package wraps this protocol: point its http-client backend at the served OpenAI-compatible endpoint and call two-step extract to get structured blocks or converted Markdown for each page image.

## Use Cases

MinerU2.5-Pro is designed for document understanding workloads at scale:

- PDF-to-Markdown conversion for LLM training data and RAG ingestion pipelines
- Table extraction from complex, rotated, or borderless layouts into HTML
- Formula recognition to LaTeX for scientific and technical documents
- Layout analysis with reading-order recovery across multi-column pages
- Chart, flowchart, and seal recognition within documents

## Performance

On OmniDocBench v1.6, MinerU2.5-Pro achieves a state-of-the-art overall score of 95.7, outperforming both specialized OCR models and much larger general-purpose frontier vision-language models, and improving substantially over the original MinerU2.5 baseline score of 92.98. Text recognition reaches an edit distance of 0.036, dense formula parsing scores 97.15 CDM, and table recognition reaches 93.62 TEDS across the benchmark, with leading results reported across five diverse table benchmarks.


Document parsing vision-language model that converts PDFs and page images to structured Markdown

MinerU2.5-Pro


### Wan2.1 VACE 14B: All-in-One Video Creation and Editing

Wan2.1 VACE 14B is an open video creation and editing model from Alibaba's Wan team that unifies multiple video tasks in a single model. Rather than shipping separate checkpoints for generation, editing, and completion, VACE (Video All-in-one Creation and Editing) accepts a text prompt together with optional video, mask, and reference-image inputs and handles the full range of tasks from that one interface. The 14B model supports both 480P and 720P output, while the smaller 1.3B variant of the family targets 480P.

#### One Model, Many Tasks

VACE covers the major video creation and editing workflows through a unified conditioning interface:

- **Text-to-Video:** Generate video directly from a text prompt
- **Reference-to-Video:** Generate video guided by one or more reference images, preserving subject identity without any preprocessing step
- **Video-to-Video Editing:** Re-render an existing video under new conditions such as depth or pose control extracted from the source
- **Masked Video Editing:** Inpaint or outpaint regions of a video using mask inputs, enabling object replacement, removal, and canvas extension
- **First-Last Frame Control:** Generate the motion between provided start and end frames
- **Task Composition:** Combine inputs, for example a reference image plus a control video, in a single generation pass

#### Built on the Wan2.1 Foundation

VACE inherits the Wan2.1 video foundation stack:

- **Wan-VAE:** A causal 3D VAE that encodes and decodes high-resolution video of arbitrary length while preserving temporal information
- **Diffusion Transformer:** The Wan2.1 DiT backbone, extended with VACE context blocks that inject video, mask, and reference conditioning
- **UMT5 Text Encoder:** Multilingual prompt understanding covering both English and Chinese, including Wan2.1's distinctive ability to render legible text in generated video

Inputs of any resolution are accepted, with best results when the source material falls near the model's native 480P and 720P generation sizes.

#### ComfyUI Integration

VACE is supported natively in ComfyUI through the WanVaceToVideo core node. This template provisions the official ComfyUI workflow templates for the model covering text-to-video, reference-to-video, video-to-video with a control video, first-last-frame generation, inpainting, and outpainting, along with the model weights repackaged by Comfy-Org at [Comfy-Org/Wan_2.1_ComfyUI_repackaged](https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged). The bundled workflows include distillation LoRAs that cut the number of denoising steps for faster iteration, and the inpainting workflow includes segmentation-assisted mask generation.

#### Use Cases

- Subject-consistent video generation from character or product reference images
- Object replacement and removal in existing footage
- Extending video canvases beyond their original framing
- Repose and restyle of existing clips via pose and depth control
- Animating between keyframes with first-last frame control
- Standard text-to-video generation

Deploy Wan2.1 VACE 14B on Vast.ai with ComfyUI to run the full creation and editing suite from a browser-based node interface.


Wan2.1 VACE is an all-in-one video creation and editing model unifying text-to-video, reference-to-video, video-to-video editing, inpainting and outpainting

Wan2.1 VACE 14B


## Overview

BGE-M3 is a multilingual text embedding model from BAAI, introduced for its versatility in what its authors call Multi-Functionality, Multi-Linguality, and Multi-Granularity. It supports more than 100 working languages and processes inputs of different granularities, from short sentences to documents of up to 8192 tokens, producing 1024-dimensional embedding vectors. Unlike earlier BGE models, BGE-M3 no longer requires adding instruction prefixes to queries.

This template serves BGE-M3's dense embeddings with vLLM through an OpenAI-compatible embeddings API. The model also carries sparse (lexical weight) and ColBERT-style multi-vector heads from its unified fine-tuning; those retrieval modes use a different serving path and are not exposed by this template.

## Key Features

- Multilingual retrieval across more than 100 languages with a single model
- Long inputs: sequences up to 8192 tokens, suited to whole-document embedding
- 1024-dimensional dense embeddings for semantic search and similarity
- Trained for all three common retrieval functionalities (dense, sparse, and multi-vector), with dense retrieval served here
- No query-side instruction prefixes needed, simplifying drop-in use in existing pipelines

## Architecture

BGE-M3 builds on the XLM-RoBERTa architecture, extended to an 8192-token maximum length via the RetroMAE pretraining recipe ([BAAI/bge-m3-retromae](https://huggingface.co/BAAI/bge-m3-retromae)), then trained with contrastive learning ([BAAI/bge-m3-unsupervised](https://huggingface.co/BAAI/bge-m3-unsupervised)) and unified fine-tuning across dense, sparse, and ColBERT objectives to produce the final [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3) checkpoint.

## Use Cases

- Semantic search and retrieval-augmented generation (RAG) over multilingual corpora
- Long-document retrieval where passages exceed the 512-token limit of most embedding models
- Cross-lingual matching, where queries and documents are in different languages
- Clustering, deduplication, and recommendation over text collections

The model card recommends a hybrid retrieval plus re-ranking pipeline for RAG: combine embedding retrieval with a lexical method such as BM25, then re-rank candidates with a cross-encoder re-ranker.

## Benchmarks

On the MIRACL multilingual retrieval benchmark, BAAI reports updated evaluation results showing BGE-M3's overall performance higher than its originally published numbers after a scoring correction. The team also released MLDR, a long-document retrieval dataset covering 13 languages ([Shitao/MLDR](https://huggingface.co/datasets/Shitao/MLDR)), on which BGE-M3's 8192-token input length is exercised. Third-party experiments cited on the model card report BGE-M3 achieving top performance in both English and other languages among the compared open and commercial embedding models.


BGE-M3 multilingual text embedding model by BAAI with 8192-token inputs and 100+ languages, served as an OpenAI-compatible embeddings API

BGE-M3


### Qwen Image Edit 2511: Instruction-Based Image Editing

Qwen Image Edit 2511 is the December 2025 release of the image editing model in the Qwen series, built on the Qwen Image foundation model. Where Qwen Image focuses on text-to-image generation, Qwen Image Edit takes one or more existing images plus a natural-language instruction and produces a precisely edited result. The 2511 release is an enhanced version over Qwen-Image-Edit-2509, with notably better consistency between the input image and the edited output.

#### What Sets 2511 Apart

The release targets the failure modes that matter most in production editing workflows. Key enhancements include mitigated image drift (unedited regions stay faithful to the source), improved character consistency, LoRA capabilities integrated directly into the base model, enhanced industrial design generation, and strengthened geometric reasoning.

#### Character and Multi-Person Consistency

Character consistency is significantly improved in this release. The model can perform imaginative edits based on an input portrait while preserving the identity and visual characteristics of the subject. Where the 2509 release improved consistency for single-subject editing, 2511 extends this to multi-person group photos, enabling high-fidelity fusion of two separate person images into a coherent group shot.

#### Multi-Image Editing

The model accepts multiple reference images in a single edit. A typical instruction references them by position, such as transferring the material from one image onto an object in another, or composing two separately photographed subjects into one scene. Reference handling is built into the conditioning path rather than bolted on, which is what makes material transfer, subject insertion, and image fusion work reliably.

#### Built-In Community LoRA Capabilities

Since the original Qwen Image Edit release, the community has developed many creative and high-quality LoRAs. Qwen Image Edit 2511 integrates selected popular LoRAs directly into the base model, unlocking their effects without extra adapters or tuning. Realistic lighting control and novel viewpoint generation, which previously required dedicated LoRAs, are now achievable out of the box.

#### Industrial Design and Geometric Reasoning

The release pays special attention to practical engineering scenarios: batch industrial product design, material replacement on industrial components, and stronger geometric reasoning, such as directly generating auxiliary construction lines for design or annotation purposes.

#### Use Cases

Qwen Image Edit 2511 fits workflows that need controlled, repeatable edits rather than open-ended generation:

- Product photography variants: material, color, and background swaps
- Character-consistent edits across marketing assets and storyboards
- Combining separately shot subjects into a single coherent scene
- Text corrections and replacements inside existing images
- Lighting adjustment and relighting of photographs
- Industrial design iteration and component material studies

#### Ecosystem

The model ships with day-one support in the ComfyUI ecosystem, including repackaged FP8 checkpoints and few-step Lightning LoRAs from the community that cut editing to as few as four sampling steps. For text-to-image generation with the same foundation model, see the Qwen Image entry.


Instruction-based image editing model in the Qwen series with multi-image inputs, strong character consistency, and precise text editing

Qwen Image Edit 2511 (FP8)


### Real-ESRGAN: Practical Image and Video Upscaling

Real-ESRGAN is a practical blind super-resolution model from the paper "Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data" (ICCVW 2021), developed by Xintao Wang and collaborators at Tencent ARC Lab. It upscales real-world photos and video frames by 4x while removing the blur, noise, ringing, and compression artifacts that accumulate in images found in the wild. Years after release it remains one of the most widely used open upscaling models, and its checkpoints power the upscaling features of many image-generation frontends.

#### Overview

Classic super-resolution models are trained on clean images downscaled with a single known operation, so they fall apart on real photographs whose degradations are unknown and compounded. Real-ESRGAN instead trains on pure synthetic data produced by a high-order degradation pipeline: multiple randomized rounds of blur, resizing, noise, and JPEG compression are applied in sequence, so the model learns to invert the messy, layered degradations of real-world images rather than a single idealized one. The result is an upscaler that behaves well on ordinary photos, screenshots, scans, and compressed video frames, not just laboratory benchmarks.

#### Key Features

-   **4x blind super-resolution** - Upscales without needing to know how the input was degraded, restoring texture and edges while suppressing noise and compression blocking
-   **Trained on synthetic high-order degradations** - Randomized chains of blur, resize, noise, and JPEG compression during training generalize to the compound degradations of real images
-   **Compact single-file checkpoint** - The RealESRGAN_x4plus generator is a small model that loads in seconds and leaves nearly all GPU memory free for large inputs
-   **Video upscaling** - Works frame-by-frame on video; the bundled workflow decodes a clip, upscales every frame, and re-encodes the result with its audio
-   **Model family** - The same codebase ships sibling checkpoints for anime imagery, anime video, and a 2x general model, all loadable through the same ComfyUI nodes

#### Use Cases

-   Upscaling AI-generated images from 1024px to 4K for print or display
-   Restoring old, low-resolution, or heavily compressed photos
-   Upscaling video clips and animation for delivery at higher resolutions
-   Cleaning up screenshots, scans, and web images before reuse
-   Batch upscaling via the ComfyUI API as a step in automated media pipelines
-   Preparing low-resolution source material for editing or compositing

#### Architecture

The generator is an RRDBNet, the deep residual-in-residual dense block network introduced by ESRGAN, trained here against a U-Net discriminator with spectral normalization for stable adversarial training on synthetic degradations. Inference in ComfyUI is tiled, so inputs far larger than the training resolution upscale within a fixed memory budget.

#### On This Template

The template runs Real-ESRGAN inside ComfyUI using only native nodes: Load Upscale Model reads the checkpoint and Upscale Image (using Model) applies it with automatic tiling. A ready-made video upscaling workflow and sample input are preloaded, and the same two nodes upscale still images by swapping the video loader for Load Image. The checkpoint served is the RealESRGAN_x4plus safetensors conversion published by the ComfyUI team at [Comfy-Org/Real-ESRGAN_repackaged](https://huggingface.co/Comfy-Org/Real-ESRGAN_repackaged).


Practical blind super-resolution model that upscales real-world images and video 4x while restoring texture and removing compression artifacts

Real-ESRGAN


### Wan2.2 S2V 14B: Audio-Driven Cinematic Video Generation

Wan2.2 S2V 14B is an open-source speech-to-video generation model from the Wan2.2 family that produces audio-synchronized, film-quality video from a reference image, an audio clip, and an optional text prompt. Rather than limiting itself to talking-head animation, S2V targets full cinematic performance: nuanced character interactions, realistic body movement, and dynamic camera work driven by the rhythm and content of the input audio.

#### How It Works

S2V conditions video generation on an audio signal. A Wav2Vec2-based audio encoder extracts features from the input speech or singing, and the diffusion model generates frames whose lip motion, facial expression, and body language follow the audio. The reference image establishes character identity and scene composition, while the text prompt steers content, style, and camera behavior. Video length adapts automatically to the duration of the input audio.

#### Key Capabilities

- **Speech- and singing-driven animation:** Characters speak or sing in sync with the provided audio track, with expression and gesture matched to delivery
- **Reference-image identity:** A single input image fixes the character, wardrobe, and setting for the generated clip
- **Text-guided direction:** Optional prompts control scene content, mood, and camera movement on top of the audio conditioning
- **Pose-driven performance:** The model additionally supports pose-sequence guidance for choreographed motion synchronized to audio
- **Long-form generation:** The underlying approach extends to longer clips and precise lip-sync editing
- **Resolution support:** Generates at 480P and 720P, with aspect ratio following the reference image

#### Wan2.2 Foundation

S2V builds on the Wan2.2 generation of video diffusion models, trained on a substantially expanded dataset compared to Wan2.1, with 65.6% more images and 83.2% more videos, plus curated aesthetic data labeled for lighting, composition, contrast, and color tone. This training recipe underpins the family's cinematic output quality and improved handling of complex motion. In benchmark evaluations reported by the Wan team, the S2V approach outperforms comparable audio-driven human animation systems, and the broader Wan2.2 family achieves superior performance against leading closed-source commercial models on Wan-Bench 2.0.

#### ComfyUI Workflow

This template deploys the model through ComfyUI using the official Wan2.2 S2V workflow. The workflow loads the FP8-quantized S2V diffusion model together with the UMT5-XXL text encoder, the Wan2.1 VAE, and the Wav2Vec2 audio encoder, and includes a distillation LoRA for fast four-step sampling. Upload an audio file and a reference image, adjust the prompt, and queue the graph to generate synchronized video. All required weights are downloaded automatically at first boot.

#### Use Cases

- Virtual presenters and spokespeople animated from a single photo and a voice track
- Music videos with characters singing in sync to a recorded performance
- Dialogue scenes and previsualization for film and animation pipelines
- Localized marketing content where one visual identity speaks many languages
- Educational and explainer videos narrated by a consistent on-screen character
- Social media content that turns portraits and voice notes into performed clips

Deploy Wan2.2 S2V 14B on Vast.ai to generate audio-driven video with ComfyUI.


Wan2.2 S2V generates audio-driven cinematic video from a reference image, an audio clip, and a text prompt

Wan2.2 S2V 14B (FP8)


### FLUX.1 [schnell]: Fast Few-Step Text-to-Image Generation

FLUX.1 [schnell] is a 12 billion parameter rectified flow transformer developed by Black Forest Labs for generating images from text descriptions. As the fastest member of the FLUX.1 family, it is trained using latent adversarial diffusion distillation, which enables it to generate high-quality images in only 1 to 4 sampling steps — a fraction of the steps required by conventional diffusion models.

#### Architecture and Design

Built on the same rectified flow transformer architecture as the rest of the FLUX.1 family, FLUX.1 [schnell] applies latent adversarial diffusion distillation to compress the generation process into a handful of steps. The result is dramatically lower latency per image while preserving the output quality and prompt adherence the FLUX.1 family is known for.

The openly released weights allow researchers and developers to study, modify, fine-tune, and build upon the model. It integrates with the reference implementation, ComfyUI for node-based local inference, and the Diffusers library.

#### Key Capabilities

FLUX.1 [schnell] combines speed with strong generation quality:

- **Few-Step Generation:** Produces complete images in 1 to 4 steps, making it well suited for interactive and high-throughput workloads
- **Output Quality:** Cutting-edge output quality and competitive prompt following, matching the performance of closed-source alternatives
- **Prompt Following:** Understands and executes complex text prompts accurately despite the compressed sampling schedule
- **Integration Flexibility:** Compatible with multiple platforms including Diffusers, ComfyUI, and various API providers

#### Use Cases

The model's speed makes it a strong fit for applications where iteration time matters:

- Rapid prototyping and ideation, where many candidate images are explored quickly
- Interactive creative tools with near-real-time feedback
- High-volume content generation for marketing and social media
- Digital art, illustration, and concept art
- Product visualization and design mockups
- A drop-in base for fine-tunes, adapters, and quantized derivatives — the model has spawned hundreds of community adapters and finetunes

#### Community and Ecosystem

FLUX.1 [schnell] has seen broad adoption since release, with monthly downloads in the hundreds of thousands. Its ecosystem includes hundreds of adapter models, dozens of finetunes and quantizations, demonstrating its role as a foundation for fast image generation applications.

#### Technical Considerations

As with other statistical image generation systems, the model is not designed to provide factual information, may occasionally fail to generate output that precisely matches prompts, and prompt-following ability varies with prompting style. Users building production systems should implement appropriate content filtering and quality validation workflows.


Fast 12B rectified flow transformer generating high-quality images in 1 to 4 steps

FLUX.1 [schnell]


### FLUX.2 [klein]: Fast Unified Image Generation and Editing

FLUX.2 [klein] is the fastest image model family from Black Forest Labs, unifying text-to-image generation and instruction-based image editing in a single compact architecture. This deployment runs the 4 billion parameter checkpoint, a rectified flow transformer distilled for end-to-end inference in as little as under a second while retaining quality that punches well above its size.

The model is built for interactive workflows, production deployments, and latency-critical applications where real-time image generation matters.

#### Key Features

- **Unified Generation and Editing:** Handles text-to-image generation, single-image editing, and multi-reference editing in one model, with no separate editing checkpoint or fine-tuning required. Generate images from text, edit an existing image with natural-language instructions, combine multiple reference images into a single composition, and keep characters, objects, and styles consistent across generations
- **Sub-Second Inference:** As the fastest distilled model in the FLUX.2 family, klein generates images in as few as 4 inference steps, making it well suited to interactive applications, rapid creative iteration, and serving pipelines where response latency is the constraint
- **Compact and Accessible:** At 4 billion parameters, klein is the most accessible model in the FLUX.2 line, designed for local development, edge deployment, and cost-efficient production serving without sacrificing output quality

#### ComfyUI Workflows

This template provisions the official ComfyUI workflows for FLUX.2 [klein]:

- Text-to-image generation, with both the distilled checkpoint and the undistilled base checkpoint available as selectable subgraphs
- Image editing with the distilled 4B checkpoint for fast instruction-based edits
- Image editing with the 4B base checkpoint for guidance-controlled edits

Example input images for the editing workflows are downloaded automatically, so every workflow runs out of the box.

#### Architecture

FLUX.2 [klein] 4B is a rectified flow transformer paired with a Qwen3 4B text encoder and the FLUX.2 VAE. Rectified flows learn a direct path between noise and image distributions, enabling high-quality sampling in very few steps. The distilled checkpoint further compresses sampling for sub-second generation, while the base checkpoint preserves classifier-free guidance for finer control and downstream fine-tuning.

#### Use Cases

- **Interactive Applications:** Sub-second generation suits user-facing products such as creative tools with live previews, conversational image assistants, and applications that generate imagery in response to user actions
- **Creative Iteration:** Fast feedback loops for concept art, illustration, storyboarding, and design exploration, with instruction-based editing to refine a composition without starting over
- **Production Content Pipelines:** Marketing assets, product visualization, social media content, and thumbnail generation at scale, where per-image latency and cost dominate
- **Multi-Reference Composition:** Combine characters, objects, and styles from multiple reference images into new scenes while maintaining visual consistency, without training LoRA adapters

Deploy FLUX.2 [klein] on Vast.ai to run official ComfyUI generation and editing workflows in minutes.


Compact Apache 2.0 rectified flow transformer for text-to-image generation and instruction-based image editing

FLUX.2 [klein]


### Z-Image-Turbo: Efficient Photorealistic Image Generation

Z-Image-Turbo is the distilled, speed-optimized member of the Z-Image family from Alibaba's Tongyi-MAI lab, released in November 2025. The family comprises 6-billion-parameter image generation models, and the Turbo variant is engineered to match or exceed leading competitors while requiring only 8 function evaluations per image — an order of magnitude fewer sampling steps than typical diffusion models. The result is sub-second inference latency on enterprise-grade GPUs without sacrificing output quality.

#### Architecture

Z-Image adopts a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture. Text tokens, visual semantic tokens, and image VAE tokens are concatenated at the sequence level into a unified input stream, maximizing parameter efficiency compared to dual-stream designs that maintain separate processing paths for text and image modalities. Text encoding is handled by a Qwen3 4B encoder, and a dedicated autoencoder handles latent-space compression.

The Turbo variant is produced from the Z-Image foundation model through a full post-training pipeline: supervised fine-tuning followed by reinforcement learning, then distillation down to 8 sampling steps with classifier-free guidance removed entirely.

#### Photorealistic Quality

Z-Image-Turbo delivers strong photorealistic generation with high aesthetic quality. The model handles fine detail in skin texture, lighting, and material rendering, making it well suited to portrait, product, and scene generation where realism matters.

#### Bilingual Text Rendering

A defining capability is accurate rendering of complex text in both English and Chinese. Rather than treating text as an afterthought, the model integrates typography into compositions coherently — a capability most open image models underserve, particularly for logographic Chinese characters.

#### Instruction Adherence and Reasoning

Z-Image-Turbo exhibits robust instruction following across compositional constraints, styles, and layouts. Paired with a prompt enhancer, the model demonstrates reasoning capabilities that go beyond surface-level descriptions, drawing on underlying world knowledge to resolve implicit details in a prompt.

#### Performance

In Elo-based human preference evaluation, Z-Image-Turbo shows highly competitive performance against leading closed models and achieves state-of-the-art results among open-source image generation models — despite running an order of magnitude fewer denoising steps than most competitors.

#### The Z-Image Family

The Turbo checkpoint sits alongside several siblings: the Z-Image foundation model focused on quality, diversity, and fine-tunability; Z-Image-Omni-Base, a raw base checkpoint for community fine-tuning covering both generation and editing; and Z-Image-Edit, a variant fine-tuned for instruction-based image editing. Turbo is the recommended choice for fast, high-quality text-to-image inference.

#### Use Cases

- Rapid photorealistic image generation where latency and cost per image matter
- Marketing and e-commerce imagery with accurate embedded text
- Bilingual content creation requiring correct Chinese and English typography
- Poster, banner, and social media asset generation
- High-throughput batch generation pipelines
- Interactive creative tools that need near-real-time feedback

#### Adoption

Z-Image-Turbo saw immediate and broad community uptake on its release, with hundreds of thousands of downloads per month on Hugging Face and day-one integration into popular inference tooling, including native ComfyUI support with official workflow templates and repackaged checkpoints.


Distilled 6B single-stream diffusion transformer that generates photorealistic images with bilingual text rendering in only 8 sampling steps

Z-Image-Turbo


### MiniCPM5 2B: Compact Open Model with Long Context and Tool Calling

MiniCPM5 2B is OpenBMB's flagship small language model: a 2.5-billion-parameter dense model built on the standard Llama architecture, with a native 131K-token context window. It targets on-device and cost-sensitive serving while claiming state-of-the-art results for its size class in reasoning, code, math, long-context and agentic tasks.

#### Key Features

- **131K native context** - long-document understanding and retrieval without context-extension tricks, backed by grouped-query attention with a very compact KV cache
- **Hybrid reasoning** - a chat-template-controlled deep-thinking mode can be switched on per request for harder reasoning problems, or left off for fast direct answers
- **Native tool calling** - the model emits structured tool calls that vLLM and SGLang convert to OpenAI-compatible responses via their built-in MiniCPM5 parsers
- **Agentic training** - post-training combines large-scale deep-thinking supervised fine-tuning, critic-based reinforcement learning, and on-policy distillation that merges multiple RL expert models
- **Standard architecture** - plain Llama-style decoder with no custom kernels or model-code forks, so mainstream inference engines load it directly

#### Architecture

The model is a deep-and-narrow dense transformer: 42 layers at a 2048 hidden size, with 16 query heads and 2 key-value heads per layer. That grouped-query layout keeps the key-value cache small enough to hold the full 131K context on a single GPU. Alongside the main checkpoint, OpenBMB publishes quantized [GGUF](https://huggingface.co/openbmb/MiniCPM5-2B-GGUF) and [GPTQ](https://huggingface.co/openbmb/MiniCPM5-2B-GPTQ) versions, a [DSpark draft model](https://huggingface.co/openbmb/MiniCPM5-2B-DSpark) for speculative decoding, and the [Base](https://huggingface.co/openbmb/MiniCPM5-2B-Base) and [SFT](https://huggingface.co/openbmb/MiniCPM5-2B-SFT) lifecycle checkpoints.

#### Benchmarks

OpenBMB reports an average of 53.9 across its evaluation suite, ahead of the listed 4B-class open models, with standout results on agentic benchmarks including 46.4 on SWE-bench Verified and 97.1 on tau-squared-Bench Telecom. The reinforcement-learning and distillation stages are credited with roughly 11 points of improvement on reasoning and general benchmarks over the SFT checkpoint.

#### Use Cases

- Agentic workflows and tool-using assistants at low cost per token
- Long-document analysis, summarization and retrieval
- Code generation and repair
- Math and structured reasoning with the thinking mode enabled
- On-device and edge deployment where memory is constrained

See the [MiniCPM5-2B model card](https://huggingface.co/openbmb/MiniCPM5-2B) for full details.


Compact 2B dense LLM with 131K context, hybrid thinking mode and native tool calling

MiniCPM5 2B


### GLM 5.3-Flash: Natively Multimodal 1M-Context MoE Model

GLM 5.3-Flash is the first natively multimodal model in the GLM-5 series from Z.ai. It has 320B total parameters with 18B active per token, accepts text, image, and video input, and serves a 1M-token context window. Z.ai reports that it outperforms GLM 5.2 across benchmarks and real-world workloads. It supports English and Chinese natively.

#### Key Features

- **Native Multimodality** - Text, image, and video inputs handled by the model itself rather than a bolted-on adapter, the first such model in the GLM-5 series
- **1M-Token Context** - Long-document analysis, large-repository navigation, and extended agentic trajectories in a single context
- **Hybrid Sparse and Linear Attention** - Combines sparse and linear attention to cut long-context serving cost while preserving long-context accuracy
- **Sparse Mixture-of-Experts** - 288 routed experts with 8 active per token, so only 18B of the 320B parameters run on any given token
- **Adjustable Thinking Budget** - A reasoning effort setting with low, high, and max levels; max is the default
- **Tool Use and Agentic Work** - Evaluated on agentic suites including Terminal-Bench 2.1, Toolathlon Verified, AutomationBench, and DeepSWE

#### Use Cases

- Visual question answering, chart and document understanding, and video comprehension
- Repository-level software engineering and code generation across very long contexts
- Multi-step agentic workflows with tool calling
- Long-document analysis, synthesis, and generation
- Terminal-based development and operations work
- Bilingual English and Chinese assistants

#### Architecture and Design

The language model has 45 layers that combine sparse attention with linear attention, alongside a 24-layer vision encoder for image and video input. Z.ai introduced Manifold-Constrained Hyper-Connections in this release to improve scaling efficiency, and the checkpoint includes a native draft layer for speculative decoding. Routing sends each token through 8 of 288 experts plus one shared expert.

#### Training and Inference

GLM 5.3-Flash starts from a newly trained base model rather than a refresh of GLM 5.2, with its architecture and training recipe redesigned around capability and efficiency, and it was pre-trained on a 30T-token multimodal corpus. Thinking is on by default. Z.ai evaluates the model at temperature 1.0 with top-p 0.95 and a maximum generation length of 163,840 tokens, and reports long-context results at context lengths up to 1M tokens.

Deploy GLM 5.3-Flash on Vast.ai for multimodal reasoning, agentic coding, and million-token context work on flexible GPU infrastructure.


Natively multimodal 320B MoE with 18B active parameters and 1M-token context

GLM 5.3-Flash


### Hy4 Preview: 770B Mixture-of-Experts Flagship with 1M-Token Context

Hy4 Preview is a new-generation Mixture-of-Experts flagship model from the Tencent Hy Team. It has 770B total parameters with 49B active per token and serves a 1M-token context window. Tencent describes it as an early release built by scaling model size, context length, and training data together, with a much larger post-training run than the previous generation, and reports the largest generation-over-generation gain the team has measured.

#### Key Features

- **Sparse Mixture-of-Experts** - 256 routed experts and 1 shared expert, with the top 8 routed experts plus the shared expert active per token, so only 49B of the 770B parameters run on any given token
- **1M-Token Context** - Long-horizon development work, large-repository navigation, and extended agentic trajectories in a single context
- **Gated DeepSeek Sparse Attention** - Sparse attention with an IndexCache that reuses the sparse index across layers, cutting long-context serving cost
- **Native Draft Layer** - A multi-token-prediction layer ships in the checkpoint for speculative decoding
- **Adjustable Reasoning Effort** - Reasoning effort defaults to high for deep chain-of-thought; a no-think setting returns direct answers
- **Domain-Trained on Expert Data** - Training data was built with Tencent software engineers, game developers, finance analysts, and security experts, and the model is co-designed with Tencent's CodeBuddy and WorkBuddy products

#### Use Cases

- Software engineering: understanding, planning, debugging, and verifying long-horizon development tasks, including front-end visual quality
- Office and analysis work: turning scattered context into documents, spreadsheets, and presentations, and building equations and financial models
- Game development: prompt-to-playable prototyping and multi-turn work against game engines
- Scientific research across AI research, molecular dynamics, condensed matter physics, and pure mathematics
- Agentic workflows with tool calling over very long contexts

#### Architecture and Design

The model has 78 layers. The first uses a dense feed-forward network and the other 77 are Mixture-of-Experts. Attention is Gated DeepSeek Sparse Attention across 64 heads, with query compression to 2048 dimensions and key-value compression to 512, and an indexer that selects the top 2048 positions. The residual pathway uses identity Hyper-Connections across four residual streams to widen information flow between layers. Hidden size is 6144 and the vocabulary is 120,832 tokens. Tencent credits DeepSeek and GLM as influences on the attention design.

#### Evaluation

Tencent ran a blind side-by-side study in which 163 internal experts rated model outputs on 203 engineering tasks. Hy4 Preview scored 2.99 against GLM 5.3 at 2.92, winning 46.8 percent of comparisons, and 2.99 against Kimi K3 at 2.94, winning 51.2 percent. Published benchmark results include 92.3 on GPQA Diamond, 85.4 on Terminal-Bench 2.1, 82.9 on SWE-bench Multilingual, 74.1 on Toolathlon Verified, 65.7 on SWE-bench Pro, 64.3 on Deep-SWE, and 62.9 on SkillsBench. Tencent labels this an early release with real headroom left, and calls out over-long reasoning on complex tasks and a tendency to over-verify its own work.

Deploy Hy4 Preview on Vast.ai for long-horizon software engineering, agentic tool use, and million-token context work on flexible GPU infrastructure.


770B MoE flagship with 49B active parameters, Gated DSA attention, and 1M-token context

MXFP8

Hy4 Preview


### Qwen3.8 Flash Next

Qwen3.8-Flash-Next is a multimodal Mixture-of-Experts model from Qwen, released as an early preview of the architecture that will underpin Qwen4. It takes text and images as input and produces text. The language model carries 125 billion parameters with 6 billion active per token, plus a 51 billion parameter N-gram embedding table and a 4 billion parameter multi-token prediction module. Context length is 262,144 tokens natively and extends to 1,000,000 tokens with YaRN.

#### Key Features

- **Hybrid attention with Qwen Sparse Attention** — three of every four layers use Gated DeltaNet, which compresses history into a fixed-size recurrent state, and the fourth uses Qwen Sparse Attention. QSA scores importance at the micro-block level rather than per token, so indexing overhead shrinks along with the attention itself. Qwen reports attention-kernel speedups of up to 10.2 times on prefill and 6.6 times on decode at one million tokens.
- **N-gram embedding** — embeddings are looked up on short n-grams of the local context rather than on a single token. This adds 51 billion parameters at almost no extra per-token compute, and because the lookup addresses are known in advance the table can be held in host memory and prefetched while the model computes.
- **Gated Residual** — the residual stream widens into four parallel branches with an element-wise data-dependent read gate and a per-branch write gate, adding expressiveness across layers while keeping training stable.
- **Ultra-sparse Mixture-of-Experts** — 512 experts with 10 routed plus one shared expert active per token, with global load balancing.
- **Built-in multi-token prediction** — a multi-step-trained MTP module whose own attention layers also use QSA, which supports speculative decoding.
- **Always reasoning** — the model always produces a thinking trace, and reasoning depth is requested per call. It also preserves thinking blocks across conversation turns by default, which helps agent workflows keep decisions consistent.

#### Benchmark Results

On coding, Qwen reports 58.7 on DeepSWE 1.1 agentic coding, 62.5 on SWE-bench Pro, 81.0 on SWE-bench Multilingual and 91.9 on LiveCodeBench v6, ahead of both Qwen3.8-27B and the much larger Qwen3.7-Plus on each. On agentic work it reports 73.9 on CoWorkBench long-horizon office tasks, 55.7 on JobBench professional job tasks and 73.5 on Toolathlon Verified real-world tool use. General reasoning scores are 91.7 on GPQA Diamond, 81.3 on IFBench instruction following and 35.9 on Humanity's Last Exam.

On vision-language work Qwen reports 84.5 on AndroidWorld mobile use, 64.0 on Vision2Web visual web development, 88.5 on RealWorldQA real-world perception, 76.6 on LVBench long video understanding and 72.3 on ERQA embodied intelligence. Visual math and chart analysis score 90.6 on MathVision and 84.6 on CharXiv reasoning questions.

Qwen states that against Qwen3.7-Plus the model cuts both training and inference cost substantially, with training taking roughly one ninth as much, while holding comparable overall quality.

#### Use Cases

- Long-horizon coding agents working across a whole repository
- Computer use and mobile use agents driving a real desktop or phone
- Tool-using assistants that call functions and read back results
- Document, chart and diagram understanding
- Long video understanding and visual question answering
- Long-context retrieval and analysis over very large inputs

#### Model Card

The full model card, including the license, is on [Hugging Face](https://huggingface.co/Qwen/Qwen3.8-Flash-Next).


180B sparse MoE vision-language model with a 51B N-gram embedding table and 262K context

Qwen3.8 Flash Next


### Muse Glimmer 30B

Muse Glimmer 30B is a multimodal reasoning model built for autonomous agentic work. It combines multi-step reasoning, reliable tool use, image and document understanding, and failure recovery in a single dense model, targeting agents that run end-to-end workflows rather than one-shot completions.

The model is published at [meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B).

#### Key Features

- **Agentic task completion** - Executes end-to-end workflows with sequential planning and recovery from failed steps.
- **Reliable tool use** - Precise, schema-based function calling that holds up across long chains of operations.
- **Multimodal reasoning** - Processes text and images together through a dedicated perception encoder, aimed at screenshots and documents.
- **Long-context understanding** - Handles contexts of 131,072 tokens and beyond.
- **Speculative decoding** - Ships with a DFlash drafter that predicts blocks of 16 tokens per forward pass using block diffusion, accelerating generation.
- **Multilingual** - Trained across more than 100 languages.

#### Architecture

Muse Glimmer 30B is a dense causal transformer of roughly 29.6B parameters across 52 layers, using a repeating pattern of local and global attention. It pairs a 6,656 hidden dimension and 32 query heads under grouped-query attention at a 16:1 ratio with sliding-window attention over 2,048 tokens and RoPE position encoding, against a 202,048-token vocabulary. Vision is handled by a ViT-G/14 perception encoder of about 1.8B parameters, integrated directly rather than bolted on as a separate adapter stage.

#### Benchmarks

On agentic coding the model reports 76.0% on SWE-Bench Verified and 51.2% on the harder SWE-Bench Pro. General agentic evaluations place it at 75.5% on MCP Atlas and 74.6% on DeepSearch QA, with 43.3% on Gaia2.

Reasoning results are led by 94.7% on AIME 2026, alongside 83.5% on GPQA Diamond and 77.0% on IFBench. Multimodal performance covers 78.8% on Charxiv Reasoning, 75.8% on OmniDocBench, and 74% on MMMU Pro.

The publisher reports that the quantized releases stay close to full precision, with degradation of 0.2% and 1.0% respectively when averaged across 15 benchmarks.

#### Use Cases

- Autonomous agents that plan a sequence of steps and invoke tools along the way
- Software engineering work, including code generation and debugging
- Multi-turn workflows built on schema-based function calling
- Interpreting visual documents, charts, and screenshots inside agentic applications
- Synthetic data generation

#### Sampling

The publisher recommends a temperature of 1.0 with top-p 0.95 and top-k 64.


Dense 30B multimodal reasoning model with 131K context

Muse Glimmer 30B


### NVIDIA Nemotron 3.5 Lightning 30B A3B BF16

NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 is the full-precision release of Nemotron 3.5 Lightning, a hybrid Mixture-of-Experts language model trained by NVIDIA. It has 30 billion total parameters with only 3 billion active per token, and was released on August 11, 2026.

Nemotron is NVIDIA's family of open models — open weights, open training data, and open recipes — aimed at building specialized AI agents efficiently. These BF16 weights are the reference checkpoint for the model: the intended starting point for post-training (supervised fine-tuning, reinforcement learning, distillation), domain adaptation, producing your own quantized or GGUF variants, and full-precision research and evaluation.

#### Architecture

The model interleaves Mamba-2 state-space layers with Mixture-of-Experts layers and a small number of Attention layers. Of its 52 layers, 23 are Mamba-2, 23 are MoE, and 6 are Attention. That hybrid design is what keeps the active parameter count at 3 billion and lets the model hold very long contexts without the memory growth a pure attention stack would incur. It supports context lengths up to 1 million tokens, with 262,144 tokens configured as the native maximum position.

The model routes to 6 of 128 experts per token, alongside a shared expert, and ships with multi-token prediction layers plus two external draft models for speculative decoding.

#### Key features

- Hybrid Mamba-2, MoE and Attention architecture with 3B active parameters out of 30B total
- Configurable reasoning: thinking can be switched on or off through the chat template, and a runtime thinking budget can be set per request
- Native tool calling, with reasoning and tool-call output parsed into separate response fields
- Long-context support, evaluated up to very long inputs
- Multilingual: English and coding languages as primary, plus Spanish, French, German, Italian and Japanese
- Pre-training data through September 2025; post-training data through May 2026

#### Benchmarks

NVIDIA measured the model under a consistent internal harness, so numbers may differ from other vendors' self-reported results.

On general knowledge it scores 81.94 on MMLU Pro. On reasoning it reaches 75.44 on GPQA Diamond without tools and 32.60 on SciCode. On coding and agentic work it scores 51.56 on SWE-bench Verified, 39.33 on SWE-bench Multilingual, 24.58 on Terminal-Bench 2.1, 85.37 on PinchBench and 36.97 on BrowseComp. On instruction following it scores 71.88 on IFBench (loose), and on long-context retrieval it scores 52.00 on AA-LCR.

Against the other models NVIDIA compared it to, Lightning 30B-A3B leads its size class on PinchBench and BrowseComp, and outperforms the larger Nemotron 3 Nano across essentially the whole suite, while trailing Qwen 3.6 35B A3B on most coding and agentic tasks.

#### Use cases

- Post-training and domain adaptation on your own data
- Producing quantized or GGUF derivatives for your own deployment targets
- Agentic applications that need tool calling and controllable reasoning
- Long-document analysis, retrieval-augmented generation and summarization
- Research and evaluation at full precision

#### Sampling

NVIDIA recommends a temperature of 1.0 and top-p of 0.95.

#### Related releases

An inference-optimized quantized release is published separately as [NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4), and the base (pre-instruction-tuning) weights as [NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16).


Hybrid Mamba-attention MoE with 3B active params, reasoning control and tool calling

NVIDIA Nemotron 3.5 Lightning 30B A3B BF16


### NVIDIA Nemotron 3.5 Lightning 30B A3B NVFP4

NVIDIA Nemotron 3.5 Lightning 30B A3B NVFP4 is the quantized, inference-ready release of Nemotron 3.5 Lightning, a hybrid Mixture-of-Experts language model trained by NVIDIA. It has 30 billion total parameters with only 3 billion active per token, supports context lengths up to 1 million tokens, and was released on August 11, 2026.

Nemotron is NVIDIA's family of open models — open weights, open training data, and open recipes — aimed at building specialized AI agents efficiently. This is the checkpoint NVIDIA recommends for serving. The weights are quantized to NVFP4, a 4-bit floating-point format, which cuts the checkpoint to roughly a third the size of the full-precision release and lets the model run on a single GPU.

#### Architecture

The model interleaves Mamba-2 state-space layers with Mixture-of-Experts layers and a small number of Attention layers. Of its 52 layers, 23 are Mamba-2, 23 are MoE, and 6 are Attention. That hybrid design is what keeps the active parameter count at 3 billion and lets the model hold very long contexts without the memory growth a pure attention stack would incur.

The model routes to 6 of 128 experts per token, alongside a shared expert, and ships with multi-token prediction layers plus separate draft models for speculative decoding.

#### Key features

- Hybrid Mamba-2, MoE and Attention architecture with 3B active parameters out of 30B total
- NVFP4 weights, quantized by NVIDIA with its own tooling rather than by a third party
- 1 million token context served by default
- Configurable reasoning: thinking can be switched on or off through the chat template, and a runtime thinking budget can be set per request
- Native tool calling, with reasoning and tool-call output parsed into separate response fields
- Multilingual: English and coding languages as primary, plus Spanish, French, German, Italian and Japanese
- Pre-trained on more than 20 trillion tokens

#### Accuracy after quantization

NVIDIA published side-by-side scores for the full-precision and NVFP4 checkpoints under a consistent internal harness, so the cost of quantization is measurable rather than assumed. The two are close across the board, and on several tasks the quantized checkpoint scores higher.

On general knowledge it scores 81.62 on MMLU Pro against 81.94 for the full-precision weights. On reasoning it reaches 75.57 on GPQA Diamond without tools, slightly ahead of full precision at 75.44. On coding and agentic work it scores 52.80 on SWE-bench Verified against 51.56, 23.46 on Terminal-Bench 2.1 against 24.58, 83.43 on PinchBench against 85.37, and 36.81 on BrowseComp against 36.97. On instruction following it scores 72.88 on IFBench (loose) against 71.88. The largest drop is on long-context retrieval, where AA-LCR falls from 52.00 to 49.19.

These numbers were measured by NVIDIA and may differ from other vendors' self-reported results.

#### Use cases

- Long-running autonomous agents and sub-agent workhorse deployments
- Agentic applications that need tool calling and controllable reasoning
- Long-document analysis, retrieval-augmented generation and summarization
- High-throughput batch serving where the smaller checkpoint leaves more room for concurrency
- Efficient inference on a single GPU

#### Sampling

NVIDIA recommends a temperature of 1.0 and top-p of 0.95.

#### Related releases

The full-precision reference weights, intended for fine-tuning and further quantization, are published separately as [NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16), and the base (pre-instruction-tuning) weights as [NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16).


NVFP4-quantized hybrid Mamba-attention MoE with 3B active params, 1M context, reasoning control and tool calling

NVIDIA Nemotron 3.5 Lightning 30B A3B NVFP4


### DeepSeek V4 Flash

DeepSeek V4 Flash is a Mixture-of-Experts (MoE) language model with 284B total parameters and 13B activated per token. This is the official release, which supersedes the earlier preview and substantially enhances agentic capability while keeping the same model structure and size. It targets highly efficient long-context intelligence, supporting a context window of up to one million tokens, and is a strong general-purpose model for reasoning, coding, and agentic workflows.

### Architecture

DeepSeek V4 Flash pairs a sparsely activated MoE design with a hybrid attention mechanism that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency. Manifold-Constrained Hyper-Connections (mHC) strengthen residual connections and stabilize signal propagation, and the model was trained on more than 32T tokens using the Muon optimizer for faster convergence. Weights use an FP4 + FP8 mixed format, with MoE expert parameters stored in FP4 and most remaining parameters in FP8.

The model exposes a reasoning effort control with three levels, low, high, and max, which set how much deliberation it spends before answering. A speculative decoding module is attached to the checkpoint for faster generation.

### Key Features

- Mixture-of-Experts model with 284B total parameters and 13B activated per token
- Context window of up to one million tokens for long-document analysis and retrieval
- Hybrid CSA and HCA attention for efficient long-context serving
- Three reasoning effort levels spanning fast responses to deep deliberation
- Native tool-calling and reasoning parsing for agentic applications
- Open-source under the MIT License

### Use Cases

- Conversational assistants and chat applications
- Coding assistance, code generation, and software engineering agents
- Mathematical problem-solving and step-by-step reasoning
- Long-context document analysis, summarization, and retrieval over very large inputs
- Agentic workflows involving tool use, browsing, and multi-step planning

### Benchmarks

The official release reports large gains on agentic and coding-agent work over the preview it replaces. It scores 82.7 on Terminal Bench 2.1, up from 61.8, and 54.4 on DeepSWE, up from 7.3. Other results include 54.2 on NL2Repo, 76.7 on Cybergym, 70.3 on Toolathlon-Verified, 25.2 on Agents' Last Exam, and 25.1 on AutomationBench Public.

DeepSeek reports that these figures exceed those of the larger DeepSeek V4 Pro preview across the published agentic benchmarks, despite Flash activating far fewer parameters per token. Code-agent evaluations were run with the max reasoning effort level.

For full model details and the complete evaluation tables, see the model card on [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731).


FP4 Mixture-of-Experts agentic model with 1M context

DeepSeek V4 Flash


### ERNIE 4.5 300B A47B

ERNIE 4.5 300B A47B is Baidu's flagship large language model, built as a Mixture-of-Experts (MoE) network with 300 billion total parameters and 47 billion activated per token. By routing each token to a small subset of specialized experts, the model delivers the quality of a very large network while keeping the compute cost of each forward pass close to that of a much smaller dense model. It is released under the Apache 2.0 license for open commercial use.

#### Architecture

The model is organized into 54 transformer layers with 64 experts, of which 8 are activated for any given token. It supports a native context length of 131,072 tokens, making it well suited to long documents, extended conversations, and large retrieval contexts. ERNIE 4.5 was developed with a scaling-efficient training stack that combines hierarchical load balancing and FP8 mixed-precision training, and its post-training pipeline applies Supervised Fine-Tuning, Direct Preference Optimization, and Unified Preference Optimization to align the model for instruction following and dialogue.

#### Key Features

-   Mixture-of-Experts design that activates only 47B of 300B parameters per token for efficient inference
-   Long 131,072-token context window for long-form and retrieval-heavy workloads
-   Supports both standard and reasoning-oriented inference styles
-   Apache 2.0 licensed for commercial deployment

#### Benchmarks

ERNIE 4.5 300B A47B posts strong results on broad knowledge and reasoning evaluations, including a score of 78.4 on MMLU-Pro, reflecting competitive performance among leading open-weight models on graduate-level, multi-domain reasoning tasks.

#### Use Cases

-   Conversational assistants and chatbots
-   Long-document summarization, analysis, and question answering
-   Content creation and writing assistance
-   Knowledge retrieval and reference-grounded generation
-   Research, reasoning, and general-purpose text generation

More detail is available on the [model card](https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT).


Baidu's 300B-parameter (47B active) Mixture-of-Experts flagship, Apache-2.0 licensed

ERNIE 4.5 300B A47B


### Laguna S 2.1 NVFP4

Laguna S 2.1 is a 117.6B total-parameter Mixture-of-Experts language model from poolside, with 8.5B parameters activated per token. It is built for agentic coding and long-horizon work, and this NVFP4 checkpoint packages the model in native 4-bit weights so it can run efficiently on modern hardware while preserving quality.

#### Overview

The model pairs a sparse Mixture-of-Experts feed-forward design (256 experts plus one shared expert) with a hybrid attention layout tuned for fast inference and low memory pressure. Of its 48 layers, 36 use Sliding Window Attention with per-head gating and 12 use global attention, in a 3:1 ratio. This mix keeps most layers operating over a short sliding window while a smaller set of global layers preserves long-range context, which reduces the KV cache footprint at long sequence lengths. The attention stack uses softplus gating with per-layer rotary scales. The KV cache is quantized to FP8 to further reduce memory per token.

#### Key features

- Mixed sliding-window and global attention for fast inference with a small KV cache
- Native reasoning support with interleaved thinking between tool calls, which can be enabled or disabled per request
- Sparse MoE routing that activates only 8.5B of 117.6B parameters per token
- A 256K-token context window, suitable for large codebases and long agent trajectories
- Native NVFP4 weights, with FP8 KV cache detected automatically at load time

#### Architecture

Laguna S 2.1 was trained with pre-training, post-training, and reinforcement-learning stages using the Muon optimizer. It has 48 layers, 256 routed experts with one shared expert, and a 512-token sliding window on its local-attention layers. The checkpoint ships configured for a 262,144-token context window, which poolside recommends for best output quality; the weights are calibrated from native 1M-token checkpoints and the context can be extended by adjusting the model configuration, with some possible quality trade-off.

#### Benchmarks

poolside reports that Laguna S 2.1, at 118B total and 8.5B active parameters, is competitive with much larger models on agentic coding and tool-use evaluations. Reported scores include 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, 59.4% on SWE-Bench Pro (public dataset), 40.4% on DeepSWE, 46.2% on SWE Atlas codebase question answering, and 49.7% on Toolathlon Verified. These results place it ahead of several models with many times its parameter count on the same tasks.

#### Use cases

- Autonomous and assisted software engineering agents
- Long-horizon coding tasks across large repositories
- Tool-calling and function-calling workflows with interleaved reasoning
- Codebase question answering and multilingual code understanding

For full model details and evaluation methodology, see the model card on [Hugging Face](https://huggingface.co/poolside/Laguna-S-2.1-NVFP4).


Agentic coding MoE (118B total, 8.5B active) with hybrid sliding-window attention, native NVFP4 weights and 256K context

Laguna S 2.1 NVFP4


### Olmo 3.1 32B Instruct

Olmo 3.1 32B Instruct is the instruction-tuned member of the [Olmo 3](https://huggingface.co/allenai/Olmo-3.1-32B-Instruct) family from Ai2 (the Allen Institute for AI). Olmo is a series of Open Language Models built to advance the science of language modeling. The Olmo 3 release is fully open: the training code, model checkpoints, datasets, and associated training details are all published, making this a strong foundation for reproducible research and transparent deployment.

This 32B dense model is derived from the Olmo-3-32B base, which is pre-trained on the Dolma 3 dataset and then post-trained on the Dolci data suite. It is designed for direct instruction following and conversational use, delivering high-quality responses across math, coding, reasoning, and general chat without requiring an explicit chain-of-thought thinking mode.

#### Key Features

- **Fully open model** - training code, checkpoints, datasets, and training details are all released by Ai2.
- **Dense 32B architecture** - a Transformer-style autoregressive language model in a single-GPU-class size for its precision.
- **Instruction tuned** - optimized for dialogue, instruction following, and assistant-style interaction.
- **Strong reasoning and coding** - competitive results on math, code, and knowledge benchmarks.
- **English language** - trained and evaluated primarily on English text.

#### Post-training Pipeline

Olmo 3.1 32B Instruct is produced through a multi-stage post-training recipe on top of the Olmo-3-32B base:

- **Supervised fine-tuning (SFT)** on the Dolci instruction datasets covering math, code, chat, and general knowledge.
- **Direct preference optimization (DPO)** to align responses with human preferences.
- **Reinforcement learning from verifiable rewards (RLVR)** to sharpen math, code, instruction-following, and chat behavior.

#### Benchmark Highlights

On the official Ai2 evaluations, Olmo 3.1 32B Instruct shows strong results across categories. In math it reaches 93.4 on MATH, 67.8 on AIME 2024, and 57.9 on AIME 2025. On reasoning it scores 84.0 on BigBenchHard and 61.7 on ZebraLogic. For coding it reaches 86.7 on HumanEvalPlus, 65.1 on MBPP+, and 54.7 on LiveCodeBench v3. On instruction following it scores 88.8 on IFEval and 39.7 on IFBench, and on knowledge and QA it reaches 80.9 on MMLU and 48.6 on GPQA. It also posts 59.8 on AlpacaEval 2 LC for chat quality and 89.5 on the safety evaluation.

#### Use Cases

- Interactive chatbots and virtual assistants
- Math, coding, and reasoning assistance
- Content creation and writing support
- Research on open, reproducible language models
- Educational and knowledge-based applications

Deploy Olmo 3.1 32B Instruct on Vast.ai to run a fully open, high-performance instruct model with vLLM or quantized llama.cpp variants.


Fully open dense 32B instruct language model by Ai2 with strong math, reasoning, and coding performance

Olmo 3.1 32B Instruct


## Overview

[Agents-A1](https://huggingface.co/InternScience/Agents-A1) is a 35B Mixture-of-Experts agentic model from InternScience, activating roughly 3B parameters per token (35B-A3B). It is built to scale heterogeneous agentic abilities across multiple domains, including long-horizon search, engineering, scientific research, instruction following, and tool-calling. The central idea behind the model is agent-horizon scaling: reaching the performance of much larger, trillion-parameter systems by scaling the horizon and diversity of an agent's behavior rather than scaling raw parameter count.

## Key Features

- **Agentic reasoning.** Agents-A1 decomposes complex tasks into executable sub-steps, plans ahead, and adapts its strategy based on intermediate results.
- **Native tool use.** The model natively supports function calling and tool integration, enabling interaction with APIs, code interpreters, search engines, and other external tools.
- **Scientific and professional reasoning.** It handles tool-integrated scientific reasoning and professional knowledge question answering.
- **Instruction following.** It precisely follows detailed, multi-constraint instructions across diverse domains.
- **Long context.** The model supports a context window of up to 262K tokens for long-horizon trajectories and extended agentic workflows.

## Architecture

Agents-A1 is a Mixture-of-Experts model with 35B total parameters and about 3B active per token. Its training investigates agent-horizon scaling from two perspectives. First, scaling long-horizon trajectories: the model is trained with the assistance of a domain-grounded knowledge-action infrastructure that jointly constructs actions, observations, and verifier outcomes, turning the agent's process into a trainable target. Second, scaling heterogeneous agent abilities through a three-stage training paradigm: full-domain supervised fine-tuning to align the base model with broad agentic behaviors, domain-level teacher models to capture specialized expertise in each domain, and multi-teacher multi-domain on-policy distillation with heterogeneity-aware optimization to improve knowledge transfer across domains.

## Use Cases

Agents-A1 targets real-world agentic and research-oriented workflows: long-horizon web search and browsing agents, engineering and coding tasks, tool-integrated scientific research, general agentic assistants that orchestrate external tools, and settings that demand precise, multi-constraint instruction following.

## Performance

InternScience evaluates Agents-A1 across six directions: long-horizon search, engineering tasks, scientific research, instruction following, general agentic tasks, and scientific agentic tasks. Despite operating in the ~35B model class, it delivers competitive results against frontier-scale systems. The model reports state-of-the-art results on several challenging benchmarks, including Seal-0 (56.4), HiPhO (46.4), FrontierScience-Olympiad (79.0), FrontierScience-Research (40.0), IFBench (80.6), and IFEval (94.8). Among comparable ~35B models it ranks best on a broad range of tasks, including BrowseComp (75.5), XBench-DS-2510 (86.0), GAIA (96.0), SciCode (44.3), HLE with tools (47.6), and MolBench-bind (56.8). Together these results position Agents-A1 as a capable and efficient agentic model that narrows the gap with much larger frontier systems while combining strong long-horizon search, robust scientific reasoning, and reliable instruction following.

For full model details, weights, and quantized variants, see the [Agents-A1 model card on Hugging Face](https://huggingface.co/InternScience/Agents-A1).


Agents-A1


### Mistral Small 4 119B

Mistral Small 4 is a powerful hybrid model that acts as both a general instruction model and a reasoning model. It unifies the capabilities of three previously separate model families - Instruct, Reasoning (previously Magistral), and Devstral (coding) - into a single, unified model. With multimodal input, an efficient architecture, and flexible mode switching, it is a strong general-purpose model for a wide range of tasks.

Compared to the previous generation, a latency-optimized setup achieves roughly a 40 percent reduction in end-to-end completion time, and a throughput-optimized setup handles about 3x more requests per second than Mistral Small 3.

#### Key Features

-   **Hybrid instruct and reasoning** - toggle between fast instant-reply mode and a reasoning mode that boosts performance with test-time compute when requested, with reasoning effort configurable per request.
-   **Vision** - accepts both text and image input and analyzes visual content alongside text, producing text output.
-   **Mixture-of-experts architecture** - 128 experts with 4 active per token, activating only a small fraction of the total parameters for each token.
-   **Long context** - supports a 256k-token context window.
-   **Agentic** - strong function-calling and JSON-output capabilities for tool use and automation.
-   **Multilingual** - supports dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Chinese, Japanese, Korean, and Arabic.
-   **Strong system-prompt adherence** - reliable behavior when steered by system prompts.

#### Use Cases

-   Developers building coding and agentic workflows for software-engineering automation and codebase exploration.
-   Enterprises deploying general chat assistants, agents, and document understanding.
-   Researchers leveraging its math and reasoning capabilities.

#### Benchmarks

According to Mistral's internal evaluations, Mistral Small 4 in reasoning mode achieves competitive scores while generating significantly shorter outputs than comparable open models. On the AA LCR benchmark it reaches a score of 0.72 using roughly 1.6K characters of output, where other open models require several times more output for comparable performance. On LiveCodeBench it outperforms a leading 120B open model while producing about 20 percent less output, reflecting its focus on both quality and speed.

Full model details are available on the [Hugging Face model card](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603).


Mistral Small 4 unifies instruct, reasoning, and coding in a single multimodal 119B MoE with vision input.

Mistral Small 4 119B 2603


### Unlimited-OCR

Unlimited-OCR is a multimodal OCR and document-parsing model from Baidu, built for one-shot long-horizon document understanding. It reads images and pages and produces structured text output, extending the ideas behind DeepSeek-OCR toward parsing long documents in a single pass.

#### Overview

The model pairs a vision front-end with a DeepSeek-style mixture-of-experts language backbone, so it can look at a rendered page and transcribe or parse it end to end. It handles both single images and multi-page inputs, including PDFs converted to page images, and returns parsed document text rather than free-form chat. Output is multilingual, making it suitable for documents in a range of scripts.

#### Key features

- Document parsing and OCR from images and multi-page inputs
- One-shot long-horizon parsing for long documents
- Single-image modes (high-detail and full-page) plus a multi-page mode for PDFs
- Multilingual text recognition
- Mixture-of-experts language backbone with a vision encoder front-end

#### Use cases

- Converting scanned documents and PDFs into structured text
- Long-document and multi-page parsing in a single pass
- Extracting text from charts, tables, and mixed-layout pages
- Building document-understanding and search pipelines over image-based archives
- Multilingual OCR across diverse scripts

#### Model resources

- Model weights and card: [baidu/Unlimited-OCR on Hugging Face](https://huggingface.co/baidu/Unlimited-OCR)


Long-document OCR and document-parsing vision-language model

Unlimited-OCR


# FLUX.1 Kontext [dev]

FLUX.1 Kontext [dev] is a 12 billion parameter rectified flow transformer from Black Forest Labs, built for in-context image editing. Unlike text-to-image models that generate from a prompt alone, Kontext takes an existing image together with a natural-language edit instruction and produces a revised image that preserves the parts you did not ask to change. It is the open-weight member of the FLUX.1 Kontext family, released to support third-party research and development.

## Overview

Kontext performs instruction-based editing: you provide a source image and describe the change you want in plain language, such as adding an object, altering a style, or adjusting a scene. The model applies the edit while maintaining consistency with the original, making it well suited to iterative workflows where an image is refined over several successive edits. It is trained using guidance distillation for more efficient inference.

## Key Features

- Change existing images based on a written edit instruction.
- Preserve character, style, and object references across edits without any finetuning.
- Robust consistency that allows an image to be refined through multiple successive edits with minimal visual drift.
- Guidance-distilled training for more efficient generation.
- Open weights intended to drive new scientific research and to empower artists to develop innovative workflows.

## Use Cases

- Targeted photo edits driven by a text instruction, for example adding, removing, or modifying elements of a scene.
- Style and appearance changes that keep the subject and composition intact.
- Reference-guided editing that carries a character, style, or object across generations.
- Multi-step creative pipelines where an image is progressively refined edit by edit.

## Architecture

FLUX.1 Kontext [dev] is a rectified flow transformer operating in latent space, applying flow matching for in-context image generation and editing. The 12B-parameter transformer is paired with text encoders and a VAE, and conditions generation jointly on the input image and the edit instruction so that outputs remain faithful to the source. A reference implementation and sampling code are provided by Black Forest Labs, and the model is available for both ComfyUI and Diffusers workflows.

## Responsible Use

The model repository includes filters for illegal or infringing content, and the FLUX.1 Kontext models were subjected to multiple rounds of pre-release and third-party safety evaluation. In its evaluations, FLUX.1 Kontext [dev] demonstrated high resilience against violative inputs relative to other similar open-weight models. Deployers are expected to keep content filters or manual review in place when using the model.

For full model details, weights, and documentation, see the model card on [Hugging Face](https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev).


12B rectified flow transformer for instruction-based in-context image editing

AI Image and Video Generation on Cloud GPUs

Why Vast.ai for AI Image and Video Generation

Popular Models for Image and Video Generation

Qwen Image 2.1

MiniMax H3

Krea 2 Turbo

Related Blogs

Related Guides

Start Building: AI Image and Video Generation on Cloud GPUs Templates

AI Image and Video Generation on Cloud GPUs

Why Vast.ai for AI Image and Video Generation

Popular Models for Image and Video Generation

.css-lw4uyf{min-height:32px;max-height:32px;min-width:32px;max-width:32px;aspect-ratio:1/1;box-sizing:content-box;-webkit-flex-shrink:0;-ms-flex-negative:0;flex-shrink:0;margin:0;color:#e9e9e9;}Qwen Image 2.1

MiniMax H3

Krea 2 Turbo

Related Blogs

Related Guides

Start Building: AI Image and Video Generation on Cloud GPUs Templates

Qwen Image 2.1