
Qwen 3.5 is Alibaba's latest model family, featuring a hybrid architecture that combines Gated DeltaNet (linear attention) with standard attention in a 3:1 ratio. This design enables dramatically faster inference — up to 8.6x at 32K context and 19x at 256K — while maintaining strong reasoning capabilities. The medium models released February 24, 2026 include:

<table>
  <thead>
    <tr>
      <th>Model</th>
      <th>Total Params</th>
      <th>Active Params</th>
      <th>Architecture</th>
      <th>VRAM (BF16)</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Qwen3.5-122B-A10B</td>
      <td>122B</td>
      <td>10B</td>
      <td>MoE</td>
      <td>~244 GB</td>
    </tr>
    <tr>
      <td>Qwen3.5-35B-A3B</td>
      <td>35B</td>
      <td>3B</td>
      <td>MoE</td>
      <td>~66 GB</td>
    </tr>
    <tr>
      <td>Qwen3.5-27B</td>
      <td>27B</td>
      <td>27B</td>
      <td>Dense</td>
      <td>~54 GB</td>
    </tr>
  </tbody>
</table>

The MoE models route each token through 8 of 256 experts plus 1 shared expert, so only a fraction of the total parameters are active per token. The 35B-A3B model activates just 3B parameters per token while having 35B total — giving you large-model quality at small-model inference cost. All three models are Apache 2.0 licensed, so no HuggingFace token is needed.

In this guide, we'll deploy Qwen3.5-35B-A3B on Vast.ai using SGLang, which has day-0 Qwen3.5 support as of v0.5.9, on a single A100 80GB GPU.

## Deploying on Vast.ai

Install the Vast.ai CLI and set your API key:

```bash
pip install --upgrade vastai
export VAST_API_KEY="YOUR_KEY_HERE"
vastai set api-key $VAST_API_KEY
```

Search for a suitable GPU. The model needs ~66 GB VRAM in BF16, so an 80 GB card works:

```bash
vastai search offers "gpu_ram>=80 num_gpus=1 direct_port_count>=1 rentable=true disk_space>=200 cuda_vers>=12.2" -o dph
```

Deploy the model using SGLang's Docker image. Replace `YOUR_INSTANCE_ID` with an offer ID from the search results:

```bash
vastai create instance YOUR_INSTANCE_ID \
    --image lmsysorg/sglang:latest \
    --env '-p 8000:8000' \
    --disk 200 \
    --onstart-cmd "python3 -m sglang.launch_server \
        --model-path Qwen/Qwen3.5-35B-A3B \
        --host 0.0.0.0 --port 8000 \
        --tp-size 1 \
        --context-length 32768 \
        --reasoning-parser qwen3 \
        --mem-fraction-static 0.85"
```

Key flags:
- `--reasoning-parser qwen3` enables the model's built-in thinking mode, separating reasoning from the final answer
- `--mem-fraction-static 0.85` allocates 85% of GPU memory for KV cache
- `--context-length 32768` sets a 32K context window (the model supports up to 262K natively)

The model will download (~70 GB) and load automatically. Check `vastai logs YOUR_INSTANCE_ID` — you'll see "The server is fired up and ready to roll!" when it's serving.

## Calling the Model

Find your instance's IP and port in the [Instances tab](https://cloud.vast.ai/instances/) of the Vast.ai console. Click the IP address button to see the port mapping for 8000/tcp.

Install the OpenAI SDK:

```bash
pip install --upgrade openai
```

Then call the model:

```python
from openai import OpenAI

VAST_IP_ADDRESS = "YOUR_IP"
VAST_PORT = "YOUR_PORT"

client = OpenAI(
    api_key="EMPTY",
    base_url=f"http://{VAST_IP_ADDRESS}:{VAST_PORT}/v1"
)

response = client.chat.completions.create(
    model="Qwen/Qwen3.5-35B-A3B",
    messages=[
        {"role": "user", "content": "What are three benefits of mixture-of-experts models? Be concise."}
    ],
    max_tokens=1024,
    temperature=0.7
)

print(response.choices[0].message.content)
```

### Response

```
1. **Computational Efficiency:** Only a subset of parameters is activated per input,
   reducing inference costs while maintaining a massive total parameter count.
2. **Enhanced Performance:** The larger effective model capacity allows for better
   accuracy on complex and diverse tasks compared to dense models of similar compute.
3. **Specialization:** Individual experts can learn to handle specific input patterns
   or domains, improving the model's robustness and adaptability.
```

With `--reasoning-parser qwen3`, the model's thinking process is automatically separated from the final answer. The thinking tokens appear in `reasoning_content` in the API response, while the clean answer appears in `content`.

## Other Models

You can swap `Qwen/Qwen3.5-35B-A3B` for the other medium models:

- **Qwen3.5-122B-A10B**: Larger MoE model, 122B total / 10B active. At BF16 the full weights need ~244 GB, so you'll need multiple GPUs (e.g., 4× A100 80GB with `--tp-size 4`). Alternatively, GGUF quantizations are available — the Q4_K_M variant is ~70 GB and fits on a single 80GB card.
- **Qwen3.5-27B**: Dense model, all 27B parameters active. Needs ~54 GB VRAM. Fits on a single A100 or H100.

## Conclusion

Qwen3.5-35B-A3B packs 35B parameters into a model that only activates 3B per token, and its hybrid DeltaNet architecture delivers substantially faster inference than standard transformers. With SGLang and a single A100 on Vast.ai, you can have it serving in under 10 minutes.

For production use, consider increasing `--context-length` — the model supports up to 262K natively, with 128K+ recommended for best reasoning performance. GGUF quantizations (Q4_K_M at ~22 GB) are also available for running on consumer GPUs like the RTX 4090.

## Useful Resources

- [Qwen3.5-35B-A3B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-35B-A3B)
- [Qwen3.5-122B-A10B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-122B-A10B)
- [Qwen3.5-27B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-27B)
- [SGLang Documentation](https://docs.sglang.ai/)
- [Vast.ai Documentation](https://docs.vast.ai/)


Deploy Qwen 3.5 medium models on Vast.ai using SGLang, including setup, launch commands, and API usage.

Running Qwen 3.5 Medium Models on Vast.ai


Liquid AI just released [LFM2-24B-A2B](https://huggingface.co/LiquidAI/LFM2-24B-A2B), a 24 billion parameter foundation model that only activates 2.3 billion parameters per token. That ratio — less than 10% of total parameters active at inference time — makes it one of the most aggressively sparse mixture-of-experts models available today.

What makes LFM2 different from other MoE models isn't just the sparsity. The architecture is a hybrid design that replaces most of the attention layers with gated short convolution blocks. Out of 40 total layers, only 10 use grouped query attention — the other 30 are convolutional. This is a fundamentally different approach from transformer-only models, and it's what allows Liquid AI to push the active parameter count so low while maintaining competitive quality across benchmarks like MMLU-Pro, GPQA Diamond, IFEval, and MATH-500.

Liquid AI, an MIT spinoff, has scaled this architecture from 350M to 24B parameters following log-linear scaling laws. LFM2 is the first model in the family large enough to be broadly useful — and with day-one support for vLLM, SGLang, and llama.cpp, it's ready to deploy today.

> **Note**: LFM2-24B-A2B is an early checkpoint trained on 17 trillion tokens without reinforcement learning or extensive post-training. A post-trained version (LFM2.5-24B-A2B) is coming.

## Model Overview

<table>
  <thead>
    <tr>
      <th>Property</th>
      <th>Value</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Developer</td>
      <td><a href="https://www.liquid.ai">Liquid AI</a></td>
    </tr>
    <tr>
      <td>Model</td>
      <td>LFM2-24B-A2B</td>
    </tr>
    <tr>
      <td>Architecture</td>
      <td>Hybrid MoE: 30 convolutional + 10 GQA attention layers, 64 experts, top-4 routing</td>
    </tr>
    <tr>
      <td>Total Parameters</td>
      <td>24B</td>
    </tr>
    <tr>
      <td>Active Parameters</td>
      <td>2.3B per token</td>
    </tr>
    <tr>
      <td>Context Length</td>
      <td>32,768 tokens</td>
    </tr>
    <tr>
      <td>Training Data</td>
      <td>17T tokens (mixed BF16/FP8)</td>
    </tr>
    <tr>
      <td>License</td>
      <td><a href="https://www.liquid.ai/lfm-open-license-v1-0">LFM Open License v1.0</a></td>
    </tr>
    <tr>
      <td>HuggingFace</td>
      <td><a href="https://huggingface.co/LiquidAI/LFM2-24B-A2B">LiquidAI/LFM2-24B-A2B</a></td>
    </tr>
  </tbody>
</table>

## Deploy LFM2-24B-A2B on Vast.ai with vLLM

At BF16, the model weights are ~48 GB — it fits on a single A100-80GB with room for a 32K context KV cache. No quantization or multi-GPU setup needed.

### Prerequisites

- A [Vast.ai](https://cloud.vast.ai) account with credits
- The Vast CLI installed and configured:

```bash
pip install --upgrade vastai  # make sure you're on v0.5.0+
vastai set api-key <YOUR_API_KEY>
```

### Find a GPU

Search for a single GPU with at least 80 GB VRAM and direct port access:

```bash
vastai search offers "gpu_ram>=80 num_gpus=1 direct_port_count>=1 cuda_vers>=12.4 cuda_vers<13" -o dph
```

> **Important**: Use a machine with **CUDA 12.x** (not 13.x). The vLLM nightly image is built against CUDA 12 and will fail with a driver mismatch on CUDA 13 machines.

### Launch the Instance

Pick an instance ID from the search results and deploy with vLLM. LFM2 requires a newer version of the `transformers` library than what ships in the vLLM image, so we upgrade it at startup:

```bash
vastai create instance <INSTANCE_ID> \
  --image vllm/vllm-openai:nightly \
  --env "-p 8000:8000" \
  --disk 100 \
  --onstart-cmd "pip install 'transformers>=5.0.0' && \
    vllm serve LiquidAI/LFM2-24B-A2B \
    --host 0.0.0.0 --port 8000 \
    --max-model-len 32768 --dtype auto \
    --trust-remote-code"
```

This upgrades the tokenizer library, pulls the model from HuggingFace, and starts an OpenAI-compatible API server on port 8000. Initial startup takes several minutes while the model downloads (~48 GB for the BF16 weights).

Once the instance is running, find your IP and port in the [Vast console](https://cloud.vast.ai/instances/) — click the IP button on your instance to see the port mapping.

> **Why `nightly` and `transformers>=5`?** LFM2's tokenizer uses the `TokenizersBackend` class introduced in `transformers 5.0`. The stable vLLM image pins `transformers<5`, so the tokenizer fails to load. Using `nightly` with a transformers upgrade resolves this. This was tested with vLLM nightly build `v0.16.0rc2`. Once a stable vLLM release ships with `transformers>=5.0.0` support, you can switch to `vllm/vllm-openai:latest` and drop the `pip install` step.

## Call the API

```bash
curl http://<VAST_IP>:<PORT>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LiquidAI/LFM2-24B-A2B",
    "messages": [{"role": "user", "content": "What makes mixture-of-experts models efficient?"}],
    "max_tokens": 300
  }'
```

## Cleanup

When you're done, destroy the instance to stop charges:

```bash
vastai destroy instance <INSTANCE_ID>
```

## Resources

- [LFM2-24B-A2B on HuggingFace](https://huggingface.co/LiquidAI/LFM2-24B-A2B)
- [LFM2-24B-A2B GGUF Quantizations](https://huggingface.co/LiquidAI/LFM2-24B-A2B-GGUF)
- [Liquid AI Blog: LFM2-24B-A2B](https://www.liquid.ai/blog/lfm2-24b-a2b)
- [Liquid AI Models](https://www.liquid.ai/models)
- [Vast.ai Documentation](https://docs.vast.ai/)
- [vLLM Documentation](https://docs.vllm.ai/)


Liquid AI released LFM2-24B-A2B, a sparse 24B MoE model with 2.3B active parameters per token.

Liquid AI's LFM2 Just Dropped — Here's How to Run It on Vast.ai


If you've ever tried generating AI video and thought, *This looks cool, but it's not what I asked for*, you're not alone!

**Wan 2.2** was designed to address scaling and efficiency challenges in video generation. It is among the first open-source video generation models to leverage a Mixture-of-Experts (MoE) architecture, enabling more efficient training. Created by Alibaba Tongyi Lab, the model gives creators more control over how their ideas turn into motion – following prompts more closely, with smoother movement and cinematic-quality results – without drastically increasing compute requirements.

On Vast.ai, Wan 2.2 is available in two variants in our [Model Library](https://vast.ai/article/vast-ai-model-library-choosing-best-ai-models):

- **[Text-to-video (T2V)](https://vast.ai/model/wan-2.2-t2v-a14b)** – Generating short videos directly from written prompts.
- **[Image-to-video (I2V)](https://vast.ai/model/wan-2.2-i2v-a14b)** – Transforming a static image into a dynamic video sequence.

Both models share the same underlying architecture, which is where things get interesting...

## What Makes Wan 2.2 Different

Most diffusion-based video models use a single neural network to handle the entire denoising process. Wan 2.2's MoE design, on the other hand, uses two specialized "experts" that strategically separate the video generation workflow:

- A **high-noise expert** that handles the early denoising stages, focusing on layout, composition, and overall scene structure and motion patterns.
- A **low-noise expert** that takes over later, refining video details and enhancing cinematic elements like lighting, contrast, and color tone.

This separation helps reduce common issues with video generation like unstable camera motion and inconsistent frames. It also makes the process more efficient. Even though each model has approximately **27 billion parameters** in total, only about **14 billion are active** per inference step.

The models automatically switch between experts via signal-to-noise ratio (SNR) thresholds – a measure of how much meaningful visual information remains versus how much noise is still present at each stage of generation. By activating the right expert at the right moment, Wan 2.2 avoids unnecessary computation. This keeps performance much more efficient, comparable to smaller single-expert models.

In short, you get more natural motion dynamics, superior visual fidelity, reduced artifacts, and less wasted compute.

## One Architecture, Two Workflows

While their underlying architecture is the same, the two Wan 2.2 variants have different creative entry points.

### Wan2.2 T2V A14B (FP8)

The text-to-video (T2V) model uses text prompts to generate **5-second clips at 480P or 720P** resolution, with granular control over lighting, composition, contrast, and color tone.

A few other capabilities of the T2V variant include:

- Superior performance on Wan-Bench 2.0 evaluations compared to leading commercial video generation systems
- Integration with Qwen models or DashScope API for enhanced prompt elaboration
- Support for parameter-type conversion, leading to improved inference speed

Overall, the text-to-video model is well suited for storyboarding, concept visualization, marketing clips, and rapid prototyping of video ideas from plain language.

### Wan2.2 I2V A14B (FP8)

The image-to-video (I2V) model begins with a static image input and turns it into a short video sequence. It offers both **480P and 720P output**, and can optionally support text prompts to guide content and motion in detail.

Some strengths of the I2V variant include:

- Handling a wide range of aesthetic preferences, from photorealistic to stylized
- Processing at **24 frames per second** (FPS) for smooth, high-definition output
- Automatic prompt derivation that makes it possible to generate videos from an image alone *without* text

The image-to-video model is especially useful for turning concept art, product photos, and illustrations into animated content.

## From Idea to Output: Designed for Real-World Use

Whether you start from a text prompt or from an image, Wan 2.2 is built for real-world creative and production workflows.

You don't even need enterprise-grade hardware to use the model. Wan 2.2's MoE architecture keeps things so efficient, you can achieve high-quality results on consumer GPUs like the RTX 4090. For larger workloads, multi-GPU inference is also an option. The model supports FSDP and DeepSpeed Ulysses for accelerated generation.

Wan 2.2 pairs particularly well with Vast.ai's flexible GPU infrastructure. On Vast, you can spin up exactly the hardware you need and run the model on a single GPU, or scale across multiple GPUs for faster generation as your video projects grow.

Both the [T2V](https://vast.ai/model/wan-2.2-t2v-a14b) and [I2V](https://vast.ai/model/wan-2.2-i2v-a14b) models are compatible with [ComfyUI](https://docs.vast.ai/documentation/serverless/comfyui-wan-2.2) workflows, making them easy to integrate into tools you may already be using – which are also available directly on Vast.ai.

Ready to start creating with Wan 2.2? Launch it on [Vast.ai](https://vast.ai/model-library) in just a few clicks and start generating cinematic video today!


Wan 2.2 uses a Mixture-of-Experts architecture to deliver efficient, high-quality AI video generation. Learn how it works and why it matters.

Wan 2.2 Explained: A New Approach to AI Video Generation


Wouldn't it be great if you could just think of a scene and instantly turn it into a video the way it appears in your head? Technology isn't quite there yet, but today’s most advanced AI video generation models are getting us closer.

Today we’re taking a look at **WAN 2.2** and **LTX-2**, two open-source/open-weights models that transform text and images into short-form video.

## What WAN 2.2 and LTX-2 Are — and How They Differ

From the outside, WAN 2.2 and LTX-2 are pretty similar tools. They’re both open-source/open-weights diffusion-based video generation models designed to turn images or text prompts into short video clips. Their underlying architecture, however, is very different.

### WAN 2.2: Prompt Fidelity and Cinematic Control

Developed by Alibaba Tongyi Lab, WAN 2.2 is built around a Mixture-of-Experts (MoE) architecture. Instead of using a single neural network to manage the entire denoising process, it employs two specialized “experts”: a **high-noise expert** for overall structure and layout, and a **low-noise expert** for refining textures and details like lighting and color tone.

Switching between these two experts means the model can allocate compute depending on what it needs to do at any given moment—focusing on broader structure first and finer details later. It also boosts efficiency by avoiding unnecessary computation.

WAN 2.2 comes in three main variants, each designed for different workflows:

- **Text-to-Video (T2V):** Generates 5-second video clips at 480P to 720P from a text prompt written in plain language. This is a flexible option for scenes where everything needs to be synthesized from scratch.
- **Image-to-Video (I2V):** Begins with a single image and turns it into a short video. It uses automatic prompt derivation to generate video from an image without text input at all, but can also support text prompts for more directed results.
- **Hybrid:** A compact model with 5 billion parameters that handles both text-to-video and image-to-video generation. It delivers high-definition results at up to 720P and 24 FPS, but is designed for users with lower VRAM.

The base WAN 2.2 models generate video only, without native audio output. However, there is a specialized speech-to-video version (WAN 2.2 S2V) that transforms static images and audio inputs into synchronized videos.

### LTX-2: Native Audio-Video Generation

Created by Lightricks, LTX-2 is a DiT-based (Diffusion Transformer) audio-video generation model. It produces audio and visuals together in one pass, keeping dialogue, lip movements, and ambient sound aligned coherently.

Its architecture is based on latent diffusion, which means the model works in a compressed version of the video first, before converting it into full resolution. This makes it more memory efficient and enables faster iteration, translating to quicker experimentation and lower hardware overhead.

LTX-2 can generate up to **~20 seconds of synchronized audio and video**, with support for high resolutions and high frame rates depending on configuration and available compute. The model offers fine-grained control options—such as LoRA-based customization and multimodal inputs including text, image, video, and audio—for precise creative direction.

This makes LTX-2 a highly flexible model. In short, it supports text-to-video, image-to-video, and native audio-visual generation, along with cross-modal workflows like audio-to-video, text-to-audio, and video-to-audio—all within a single model.

## Choosing the Right Model for Your Workflow

How the two models are designed directly affects what you experience as a user.

For instance, WAN 2.2's MoE design prioritizes structured generation and motion consistency. It boasts strong prompt adherence with high-fidelity output and is more likely to preserve scene intent across frames, sticking closely to what you asked for—albeit at the cost of slightly longer generation times.

LTX-2's latent diffusion approach emphasizes speed and accessibility. It’s faster to iterate with, easier to experiment on, and even offers native audio-video sync. However, it may require more prompt tuning to get exactly what you want.

### Choose **WAN 2.2** if you want:

- Cinematic or narrative-style clips where composition and camera motion are critical
- Strong prompt fidelity for complex scenes with multiple elements
- More deliberate, production-oriented outputs and professional video content

### Choose **LTX-2** if you prefer:

- Rapid prototyping of video concepts and creative exploration for lengthier scenes
- Visual storytelling or character-driven video with synchronized dialogue or sound
- A lighter and more iterative workflow where speed matters more than precision

Both models also integrate with ComfyUI, so you can jump right into testing them out with an intuitive node-based visual workflow.

## Final Thoughts

Neither WAN 2.2 nor LTX-2 is objectively superior to the other. The two open-source/open-weights models are designed for different kinds of workflows and creative goals. The best way to get a feel for them is to actually try them out.

The good news is that both models run well on high-end consumer GPUs, making them far more accessible than many people might expect. With Vast.ai, it’s even easier: you can spin up the right hardware on demand and experiment on your own terms, paying only for the compute you need—and save up to 80% over traditional clouds.

Try [WAN 2.2 T2V](https://vast.ai/model/wan-2.2-t2v-a14b) and [WAN 2.2 I2V](https://vast.ai/model/wan-2.2-i2v-a14b), or [LTX-2](https://vast.ai/model/ltx-2) (or both!) in our [Model Library](https://vast.ai/model-library), and build your own creative pipeline on Vast.ai today.


A practical comparison of WAN 2.2 and LTX-2—two open-source/open-weights AI video models—covering architecture, strengths, and use cases.

Posts about: Models

Running Qwen 3.5 Medium Models on Vast.ai

Liquid AI's LFM2 Just Dropped — Here's How to Run It on Vast.ai

Wan 2.2 Explained: A New Approach to AI Video Generation

WAN 2.2 vs. LTX-2: Which AI Video Model Should You Use?