Atria Dawn: Shanghai AI Lab's 744B open agent lands, MIT-licensed
A 256K MoE built for long-horizon research: strong at browsing and search, weaker at code, with open weights and a free OpenAI-compatible preview API.
Copy markdown
What dropped: a 744B open agent under MIT
Shanghai AI Lab (InternLM) posted Atria Dawn Preview — a 744B mixture-of-experts model with a 256K context window and an MIT license, in both full and FP8 checkpoints. The weights are on Hugging Face now; the arXiv paper landed Sept 15.
Where it's strong: search and browsing
It's tuned for long-horizon research agents, and the self-reported scores lean that way: DeepSearchQA 96.0 and BrowseComp 92.5. If your agent's job is to dig through the web and tools across many steps, that's the pitch.
Where it isn't: coding
Independent tracking is cooler. benchlm ranks it #31 of 153 on agentic tasks, and SWE-bench Pro comes in at 59.6 against Claude Fable 5.1's 81.2. Don't swap out your coding model — reach for this when the task is research, not refactors.
Run it today: free API or your own GPUs
There's a free, OpenAI-compatible preview endpoint at api.atria-asi.ai with a 262K window — point an existing client at it and go. To self-host, the weights run on vLLM 0.23+ or SGLang 0.5.13+, though a 744B MoE still wants serious hardware.
Build this weekend: a verifiable research agent
The model is trained around executable, reproducible tool use, making it a natural backend for a deep-research loop: browse, run code, check the result, repeat. The free API means you can prototype one before renting a single GPU.
The catch: preview, unstable, self-scored
Every headline benchmark is vendor-reported and unverified, the preview API is flagged as unstable, and structured output is reportedly broken through LLMGateway. Treat it as an experiment to wire up, not a production dependency yet.