OmniAtlas-Qwen2.5-3B

<h1 style="text-align: left; font-size: 1.6em; margin-bottom: 0.75em;"> <span style="color:#1628a7; font-weight:bold;">O</span><span style="color:#402b94; font-weight:bold;">m</span><span style="color:#673ea0; font-weight:bold;">n</span><span style="color:#8b16aa; font-weight:bold;">i</span>Atlas: Omni-Modal Foundation Agent </h1> <div style="text-align: left; margin-bottom: 18px;"> <a href="https://arxiv.org/abs/2602.22897" target="_blank">📄 Paper</a> &nbsp; • &nbsp; <a href="https://github.com/RUC-NLPIR/OmniGAIA" target="_blank">💻 Code & Demo</a> &nbsp; • &nbsp; <a href="https://huggingface.co/collections/RUC-NLPIR/omnigaia" target="_blank">🤗 Dataset & Model</a> &nbsp; • &nbsp; <a href="https://huggingface.co/spaces/RUC-NLPIR/OmniGAIA-Leaderboard" target="_blank">📈 Leaderboard</a> </div>

This is the official checkpoint for the OmniAtlas-Qwen2.5-3B model, introduced in the paper OmniGAIA: Towards Native Omni-Modal AI Agents. Download and use it for evaluation on the OmniGAIA Benchmark, or apply it to your own omni-modal agentic tasks.

OmniAtlas Training Pipeline

<div align="center"> <img src="./assets/omniatlas_training.png" width="95%" /> </div>

OmniAtlas is trained in two stages:

1. Trajectory Synthesis & Supervised Learning — Gemini-3 provides step supervision while DeepSeek-V3.2 performs tool-augmented reasoning. Successful trajectories are used for SFT.
2. OmniDPO: Fine-Grained Error Correction — Gemini-3 identifies and corrects errors in failed trajectories across perception, reasoning, and tool-use dimensions, producing preference pairs for DPO training.

Example Usage with vLLM

Serve the model with vLLM via an OpenAI-compatible API:

vllm serve /path/to/your/OmniAtlas-Qwen2.5-3B \
    --served-model-name omniatlas-qwen2.5-3b \
    --port 8080 \
    --host 0.0.0.0 \
    --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.9 \
    --trust-remote-code \
    --enable-auto-tool-choice \
    --tool-call-parser hermes \
    --uvicorn-log-level debug \
    --max-model-len 32768

Then query the model with omni-modal inputs (image, audio, video) and tools:

import base64
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="empty")

def encode_base64(file_path):
    with open(file_path, "rb") as f:
        return base64.b64encode(f.read()).decode("utf-8")

video_b64 = encode_base64("/path/to/video.mp4")
audio_b64 = encode_base64("/path/to/audio.wav")
image_b64 = encode_base64("/path/to/image.jpg")

tools = [
    {
        "type": "function",
        "function": {
            "name": "web_search",
            "description": "Perform a Google web search and return the top results (title, URL, snippet).",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string", "description": "The search query string."}
                },
                "required": ["query"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "page_browser",
            "description": "Fetch and extract the full text content from one or more webpages.",
            "parameters": {
                "type": "object",
                "properties": {
                    "urls": {
                        "type": "array",
                        "items": {"type": "string"},
                        "description": "A list of webpage URLs to fetch content from."
                    }
                },
                "required": ["urls"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "code_executor",
            "description": "Execute Python code in a sandbox. Available: numpy, pandas, sympy, math, etc.",
            "parameters": {
                "type": "object",
                "properties": {
                    "code": {"type": "string", "description": "Python code to execute."}
                },
                "required": ["code"]
            }
        }
    },
]

response = client.chat.completions.create(
    model="omniatlas-qwen2.5-3b",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "video_url", "video_url": {"url": f"data:video/mp4;base64,{video_b64}"}},
                {"type": "audio_url", "audio_url": {"url": f"data:audio/wav;base64,{audio_b64}"}},
                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_b64}"}},
                {"type": "text", "text": "Based on the video, audio, and image, answer the question: ..."},
            ],
        }
    ],
    tools=tools,
    tool_choice="auto",
)
print(response.choices[0].message.content)

For full evaluation and agent usage, see the OmniGAIA repository.

Citation

If you find OmniAtlas useful in your work, we kindly ask that you cite us:

@misc{li2026omnigaia,
      title={OmniGAIA: Towards Native Omni-Modal AI Agents}, 
      author={Xiaoxi Li and Wenxiang Jiao and Jiarui Jin and Shijian Wang and Guanting Dong and Jiajie Jin and Hao Wang and Yinuo Wang and Ji-Rong Wen and Yuan Lu and Zhicheng Dou},
      year={2026},
      eprint={2602.22897},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.22897}, 
}