
DR-Tulu is AI2's open-source research agent—an 8B model that autonomously plans research strategies, searches the web, reads pages, and synthesizes comprehensive answers with citations. Unlike typical LLMs with tools bolted on, DR-Tulu was trained end-to-end with its MCP tools, making it a compelling open alternative to proprietary research APIs. This guide walks through deploying DR-Tulu on Vast.ai, where you can run the full agent stack on affordable GPU instances and integrate it into your own agentic applications.

## What Makes DR-Tulu Special?

Key features include:
- **Native MCP Integration**: Trained with web search and page reading tools from the start
- **Autonomous Research**: Plans strategies, gathers sources, and synthesizes cited answers
- **Efficient Architecture**: 8B parameters with strong research capabilities

## What's in the New Guide

Our latest documentation walks you through deploying DR-Tulu on Vast.ai with a split architecture: vLLM for GPU inference on Vast.ai, MCP backend running locally for tool execution.

**Complete Deployment Instructions**
- Instance selection with the Vast.ai CLI
- vLLM configuration with the correct version (critical for compatibility)
- MCP backend setup with required API keys
- Workflow configuration to connect the components

**Three Ways to Use DR-Tulu**
- Interactive chat for exploratory research
- Batch evaluation against built-in datasets
- Python API for programmatic integration

## Why This Architecture?

The split deployment keeps GPU-intensive inference on Vast.ai while running tool orchestration locally. This means:

- Your API keys (Serper, Jina) stay on your machine (or wherever you deploy the MCP backend)
- Model inference doesn't require sending data to third-party APIs
- You can modify the MCP backend without redeploying the model
- Lower Vast.ai costs since you only need GPU for inference

## Who Should Use This Guide?

This deployment guide is perfect for:
- **Researchers** who need cited, source-backed answers at scale
- **Developers** building applications that require web research capabilities
- **Teams** evaluating open alternatives to proprietary research APIs

## Get Started

The complete guide is now available in our documentation:

**[Read: Running DR-Tulu on Vast.ai →](https://docs.vast.ai/examples/mcp/dr-tulu)**

Whether you're exploring agentic AI architectures or ready to deploy your own research assistant, this guide provides everything you need to get DR-Tulu running on Vast.ai infrastructure.

Ready to try it? [Sign up for Vast.ai](https://vast.ai) and follow the guide to deploy your first instance.


Learn how to deploy DR-Tulu, AI2's open-source research agent, on Vast.ai for autonomous web research with citations.

Deploy Deep Search with DR-Tulu on Vast.ai


dstack is an open-source GPU orchestration platform that automates instance provisioning and lifecycle management across cloud providers. This guide walks through using dstack with Vast.ai as the backend, combining declarative infrastructure with competitive GPU marketplace pricing.

## What Makes dstack Special?

Key features include:
- **Infrastructure as Code**: Define GPU requirements, pricing limits, and workloads in YAML
- **Automatic Provisioning**: dstack finds and provisions the best available instances
- **Cost Controls**: Set `max_price` to cap hourly costs automatically
- **Built-in Proxy**: Access services through dstack's authenticated endpoint

## What's in the New Guide

Our latest documentation walks you through deploying language models with dstack and vLLM on Vast.ai.

**Complete Setup Instructions**
- Installing and configuring dstack with your Vast.ai API key
- Starting the dstack server and CLI
- Creating service configurations for vLLM deployments

**Working Service Configurations**
- Ready-to-use YAML for deploying Qwen3-30B-A3B
- GPU memory and pricing parameters
- Real deployment outputs showing what to expect

**API Integration Examples**
- Python code using the OpenAI SDK
- cURL examples for testing
- Streaming response implementations

## Why dstack + Vast.ai?

The combination gives you orchestration on top of a GPU marketplace:

- **Simplified Workflow**: No manual instance hunting or environment setup
- **Cost Optimization**: dstack finds the cheapest instance meeting your requirements
- **Flexible Pricing**: Access Vast.ai's competitive rates with automatic cost caps
- **Production-Ready API**: vLLM provides OpenAI-compatible endpoints

The guide demonstrates deploying Qwen3-30B-A3B on an H100 80GB—provisioned with a single command.

## Who Should Use This Guide?

This deployment guide is perfect for:
- **Teams** wanting reproducible, version-controlled GPU deployments
- **Developers** building LLM applications who want simple infrastructure management
- **Anyone** tired of manually provisioning and configuring GPU instances

## Get Started

The complete guide is now available in our documentation:

**[Read: Deploy LLMs with dstack and vLLM on Vast.ai →](https://docs.vast.ai/examples/serving-infrastructure/dstack-vllm)**

Whether you're new to GPU orchestration or looking for a better way to manage LLM deployments, this guide provides everything you need to get started with dstack on Vast.ai.

Ready to try it? [Sign up for Vast.ai](https://vast.ai) and follow the guide to deploy your first model.


Learn how to deploy language models with dstack and vLLM on Vast.ai, combining declarative infrastructure with competitive GPU marketplace pricing.

Deploy LLMs with dstack on Vast.ai


As 2025 comes to a close, we're looking back at the milestones and updates that shaped the past year for us here at Vast.ai.

What a year it's been! From new hardware to major security achievements and product launches, here are our top moments of 2025.

## Expanding Our Global GPU Fleet

The past year saw some of the biggest hardware expansions in Vast.ai's history.

In 2025, we added NVIDIA's latest RTX 5000-series GPUs – including the RTX 5090 – followed by high-performance machines like the RTX PRO 6000 WS and the H200 and B200. These additions supported higher-throughput workloads and more efficient AI inference across a [wide range](https://cloud.vast.ai/) of hardware profiles.

Another highlight: our RTX 5090 inventory surpassed 1,000 machines this year!

## Security and Compliance Milestones

This year, Vast.ai achieved both **SOC 2 Type I** and **Type II** certification. These independent audits validated that the security infrastructure we've built over 6+ years of trusted service continues to meet rigorous compliance standards.

We take this responsibility seriously. Data privacy and integrity remain core priorities in everything we build. As we continue to advance our [security and compliance](https://vast.ai/compliance) roadmap, we're committed to upholding the trust you place in us.

## New Templates and Tools

We introduced a number of ready-to-use templates this year, to help you move from idea to execution quickly and seamlessly. Some notable ones included:

- [Langflow + Ollama](https://docs.vast.ai/langflow-ollama) for visual AI workflow design
- [Ostris AI Toolkit](https://cloud.vast.ai/?ref_id=62897&template_id=81e6e9dd266d9bdbcbb0fd48b9afe357) for custom Stable Diffusion and LoRA training
- [Open-Sora](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=Open-Sora) and [LTX Video](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=ComfyUI%2520%252B%2520LTX%2520Video) (ComfyUI) for video generation workflows
- [ComfyUI ACE Step](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=ACE%2520Step%2520Music%2520Generation) for AI-assisted music creation
- [InvokeAI](https://cloud.vast.ai/?ref_id=62897&template_id=c309df384010cf9d3dc9abb9e0a5a6fd) for image generation in a professional web interface
- [Dia 1.6b TTS](https://cloud.vast.ai/?ref_id=62897&template_id=6ba071acb5f590913c5eaf6c2dc3d0db) for lifelike text-to-speech

But the biggest update of the year was still ahead...

## Introducing Vast.ai Serverless

The headline story of 2025 was the [recent launch](https://vast.ai/article/vast-ai-serverless-automated-gpu-scaling) of **Vast.ai Serverless** – a fully automated GPU compute layer designed for AI inference at scale.

With our Serverless offering, you can easily run workloads without managing instances on your own or overpaying for capacity. You simply send inference requests and the platform automatically selects the right hardware, routes workloads across regions as needed, and predictively scales up and down based on usage.

In short, Vast.ai Serverless gives you access to a diverse, globally distributed GPU fleet with a serverless model that applies predictive optimization and flexible scaling – with none of the overhead. We're thrilled to finally make truly serverless AI inference a reality on Vast.ai!

## Looking Ahead

To all of our users and the wider Vast.ai community, thank you for being part of our journey this year. It's been an incredible year of growth, and we're grateful to everyone building and innovating with us.

Here's to an even stronger year ahead!


As 2025 comes to a close, we're looking back at the milestones and updates that shaped the past year for us here at Vast.ai.

Vast.ai's 2025 Year in Review


With so many AI model libraries to choose from, making the right decision requires research and testing. Luckily, [Vast.ai](http://vast.ai) partners with some of the leading innovators across a variety of generative and assistive models. Along with our GPU rentals, we empower businesses and individuals to produce their best work without latency issues or breaking the bank. Not only are these AI models reliable, they are also very affordable. You can scale production and workload with ease. Whether looking for an image generator, video production, audio, text or computer vision research, these are some of the best AI models available.

Model Library is Vast's new, model-first way to deploy AI. Instead of starting from the runtime-level (like vLLM, ComfyUI, or PyTorch), you can browse and discover model-level templates, sorted by modality and use case, then launch a running instance in seconds with full custom control over the underlying runtime, parameters, and infrastructure.

Every template includes recommended hardware presets and clear dedicated GPU rental pricing, so you can quickly get started with the right cost/performance setup and deploy with confidence, or tweak to fit your needs.

Coming soon: more runtimes, more model variations, and Vast Serverless model templates.

## The Best AI Image Generation Models

With text-based prompts, you can produce images from a range of popular AI models. From hyper-realistic renderings to cartoonish depictions, these models are among your best bets for generating images.

- **[HiDream I1 Full](https://vast.ai/model/hidream-i1-full)**: Released in May 2025, HiDream produces images in a variety of styles. Whether you want realistic photo imagery or stylized cartoon graphics, HiDream can deliver. With speed and precision, it can produce images without sacrificing quality.

- **[Qwen Image (FP8)](https://vast.ai/model/qwen-image)**: With a specific focus on generating images containing text, Qwen Image can produce high-quality images that provide a strong eye for detail. This is true not only of images with English characters, but even intricate characters for multilingual projects. It has a particular focus on generating images containing Chinese characters.

- **[FLUX.1 [dev]](https://vast.ai/model/flux.1-dev)**: Whether prototyping concepts, marketing, or digital art, FLUX.1 is an open-source tool that rivals many closed-source competitors in its functionality and precision. Developed by Black Forest Labs, this 12-billion-parameter image generator serves for a range of use cases from concepting to marketing materials and scientific visualization.

- **[Juggernaut XI v11](https://vast.ai/model/juggernaut-xi-v11)**: With a particular emphasis on photorealistic images, Juggernaut XI v11 is trained to provide exceptional detail on challenging human features like hands and facial expressions. It also provides exceptional detail for landscapes and other detailed imagery that have historically fallen into the realm of "uncanny valley."

- **[RealVisXL V5.0](https://vast.ai/model/realvisxl-v5.0)**: Developed by Evgeny, RealVisXL renders photorealistic images and produces anatomically accurate and high-fidelity visuals thanks to inference parameters like DPM++ SDE Karras sampling. This makes it equally adept at producing architectural designs and photorealistic images.

- **[Stable Diffusion XL Base 1.0](https://vast.ai/model/stable-diffusion-xl-base-1.0)**: Developed by Stability AI, Stable Diffusion XL Base improves image quality by providing two models. With a brainstorming mode in its base model, you can parse through a variety of images and then hone in on specific imagery with the more detailed refinement module. This holistic approach provides you with everything you need to craft the specific images you need.

## The Best AI Video Generation Models

[Vast.ai](http://vast.ai) offers some of the best video generation options available on the market. With options from a variety of AI engines and developers, you can produce stunning videos that rival cost-and-labor-intensive CGI and video productions.

- **[LTX Video](https://vast.ai/model/ltx-video)**: The first Diffusion Transformer (DiT)-based video generation model, LTX Video, produces high quality videos at resolutions and framerates that exceed much of the competition. Videos are available at 30 FPS and a 1216×704 resolution. This gives it an advantage over many of its competitors.

- **[Mochi 1 Preview](https://vast.ai/model/mochi-1-preview)**: Specializing in photorealistic video generation, Mochi 1 has 10 billion parameters that are built on an innovative Asymmetric Diffusion Transformer architecture. This helps it process over 44,000 visual tokens to produce high-quality videos.

- **[Wan2.2 I2V A14B (FP8)](https://vast.ai/model/wan-2.2-i2v-a14b)**: Wan2.2 I2V generates 720p videos from static image prompts. With a high-noise expert option for background imagery and a low-noise expert option for movement, you can make still images come to life with striking detail.

- **[Wan2.2 T2V A14B (FP8)](https://vast.ai/model/wan-2.2-t2v-a14b)**: Released in July 2025, Wan2.2 T2V generates 5-second videos at both 480P and 720P resolutions. With cinematic aesthetics and complex motion capabilities, it provides greater detail than many previous open-source and commercial models.

## The Best Audio Generation AI Models

Whether you need to produce a podcast, narrate an audiobook, engineer music production, or just explore the possibilities of ai-generated sounds, [Vast.AI](http://vast.ai) provides affordable audio generation options from some of the best-known models available.

- **[ACE Step V1 3.5B](https://vast.ai/model/ace-step-v1-3.5b)**: ACE Step V1 produces realistic audio thanks to its 3.5-billion-parameter model. It can be leveraged for a range of use cases: podcasts, audiobooks, and a wide range of music production prompts.

- **[Dia 1.6B](https://vast.ai/model/dia-1.6b)**: Developed by Nari Labs, Dia 1.6B is an audio generation model that specializes in multi-speaker production. This works great for podcasts and other text-to-audio prompts and excels at inflecting speakers' voices and producing sighs, laughter, and other non-verbal flourishes.

## The Best Text Generation AI Models Available

Whether you need agentic support, assistance on scientific research, or multilingual text translation, there are many incredible choices available. These are the best text generation AI models available from [Vast.AI](http://vast.ai).

- **[GLM 4.6](https://vast.ai/model/glm-4.6)**: This model provides support for language models that work on multi-step workflows. With over 200,000 tokens and sophisticated reasoning logic, it can excel at anything from coding applications to agentic capabilities.

- **[Kimi K2 Thinking](https://vast.ai/model/kimi-k2-thinking)**: Released under a Modified MIT License, Kimi K2 Thinking supports both commercial and research applications. Its chain-of-thought reasoning allows it to operate through hundreds of steps without degrading its response quality.

- **[DeepSeek V3.2 Exp](https://vast.ai/model/deepseek-v3.2-exp)**: Built for complex research and debugging scenarios, DeepSeek V3.2 Exp is an incredible assistant for very labor-intensive processes. With the ability to parse and analyze large bodies of text, you can automate some of the most tedious parts of your day in seconds.

- **[Kimi K2 Instruct 0905](https://vast.ai/model/kimi-k2-instruct-0905)**: With over a trillion parameters, Kimi K2 Instruct excels at language models including agentic intelligence and software development. Its massive 256k token window provides advanced decision-making capabilities for a range of issues and use cases.

- **[Qwen3 235B A22B Thinking 2507](https://vast.ai/model/qwen3-235b-a22b-thinking-2507)**: Developed by Alibaba, this 235-billion-parameter Mixture-of-Experts (MoE) language model is specifically engineered for extended reasoning tasks by activating 8 experts and 22 billion parameters per token. This makes it ideal for complex projects including mathematics, scientific research, and code generation.

- **[Qwen3 Coder 480B A35B Instruct](https://vast.ai/model/qwen3-coder-480b-a35b-instruct)**: This 480-billion-parameter Mixture-of-Experts (MoE) model developed by Alibaba is designed for advanced code generation and is optimized for actionable code output by operating exclusively in an efficient non-thinking mode.

- **[DeepSeek R1 0528](https://vast.ai/model/deepseek-r1-0528)**: This advanced reasoning language model developed by DeepSeek AI utilizes advanced chain-of-thought logic for complex mathematical and logical problems. This makes it a great choice for tool and API integrations.

- **[DeepSeek V3.1](https://vast.ai/model/deepseek-v3.1)**: This hybrid language model operates in both thinking and non-thinking modes. This dual-mode architecture is equally suited for deep reasoning with visible thought processes or generating fast responses without intermediate reasoning.

- **[GPT-OSS-120b](https://vast.ai/model/gpt-oss-120b)**: GPT-OSS-120b is an open-weight reasoning model from OpenAI designed for production use cases. With a range of low, medium, and high reasoning settings it balances speed and accuracy for agentic functions.

- **[GPT OSS 20b](https://vast.ai/model/gpt-oss-20b)**: Built for lower latency and specialized use cases. For everything from web browsing to function calling and API integration tasks, GPT OSS 20b is a great choice for agentic support.

- **[Llama 4 Maverick 17B 128E Instruct](https://vast.ai/model/llama-4-maverick-17b-128e-instruct)**: With seamless processing of both text and visual inputs, Llama 4 Maverick 17B supports up to a 10 million token context length and can process up to 5 input images simultaneously. It supports 12 languages for a wide array of agentic functions including English, Spanish, Tagalog, French, German, Hindi, Indonesian, Italian, Portuguese, Arabic, Thai, and Vietnamese.

- **[Llama 4 Scout 17B 16E Instruct](https://vast.ai/model/llama-4-scout-17b-16e-instruct)**: Released in April 2025, Scout 17B 16E Instruct employs a mixture-of-experts (MoE) and has a total of 109 billion parameters. It functions as a great assistant for analyzing data, parsing large quantities of data and is compatible with 12 languages including English, Spanish, Tagalog, French, German, Hindi, Indonesian, Italian, Portuguese, Arabic, Thai, and Vietnamese.

- **[Qwen3 VL 235B A22B Instruct](https://vast.ai/model/qwen3-vl-235b-a22b-instruct)**: The most powerful vision-language model in the Qwen series, VL 235B A22B Instruct combines 236 billion parameters across visual understanding, agent applications, context processing, and multimodal reasoning tasks. It excels at recognizing a range of items from landmarks, to plants and animals, to works of art and celebrities. This makes it a great research tool for imaging, robotics, and even technical documentation.

## The Best Computer Vision Models Available

When it comes to choosing an AI model that can interpret visual elements from real-world scenarios, these are among the best:

- **[DeepSeek OCR](https://vast.ai/model/deepseek-ocr)**: This vision-language model specializes in optical character recognition and document understanding. This makes it great for extracting text from images or scanned documents, processing PDFs, and much more.

- **[GLM 4.5V](https://vast.ai/model/glm-4.5v)**: Whether summarizing longform videos, analyzing data, or scanning documents for text, GLM 4.5V provides support for advanced reasoning tasks. With so many use cases, you can even toggle between quick responses or deeper analysis of data to make it work for you the way you want.

- **[InternVL3 78B](https://vast.ai/model/internvl3-78b)**: Whether analyzing technical documentation, interpreting results from medical images, or providing agentic support, InternVL3 78B has many use cases from text-only applications to a mixture of visual applications.

- **[Llama 4 Maverick 17B 128E Instruct](https://vast.ai/model/llama-4-maverick-17b-128e-instruct)**: Featuring 17 billion activated parameters distributed across 128 total experts, this model combines text and image understanding capabilities within a unified architecture. This model provides everything from an assistant-like conversational experience to image captioning, document analysis, and multilingual support.

- **[Llama 4 Scout 17B 16E Instruct](https://vast.ai/model/llama-4-scout-17b-16e-instruct)**: With efficiency being front and center, this model runs much more leanly than Maverick. This gives it a more conversational tone while still providing agentic support, robust analysis of charts, multilingual functionality.

- **[Qwen3 VL 235B A22B Instruct](https://vast.ai/model/qwen3-vl-235b-a22b-instruct)**: Qwen3's most powerful vision-language model, it provides incredible PC and mobile GUI navigation for automation tasks, visual coding support across a range of coding languages, and agentic support.

---

If you're ready to take your AI workflows to the next level, [Vast.ai](http://vast.ai) is here to help. From an [AI model library](https://vast.ai/model-library) with a host of tools and use cases to [GPU servers](https://cloud.vast.ai/) that can accelerate your research, we're here to help empower and scale your discovery every step of the way.

---


Explore Vast.AI's comprehensive model library featuring the best AI models across image generation, video production, audio, text, and computer vision.

Deploy Deep Search with DR-Tulu on Vast.ai

What Makes DR-Tulu Special?

What's in the New Guide

Why This Architecture?

Who Should Use This Guide?

Get Started

Deploy LLMs with dstack on Vast.ai

Vast.ai's 2025 Year in Review

Vast.AI's Model Library: Choosing from the Best AI Models Available

Subscribe for our product updates.