Shanghai AI Lab ships a 744B agentic model with no announcement
The lab that spent last week talking about science as the next programming just put a frontier-class agent on Hugging Face without a press release — and the weights, not the blog post, are the argument.
Shanghai AI Laboratory has posted Atria Dawn Preview, a 744-billion-parameter Mixture-of-Experts agentic model, under the MIT license on Hugging Face and ModelScope — with no announcement, no launch post, and no paper. The model card is the only statement the lab has made: it describes a "preview version of a new-generation agentic model" built on the 744B MoE GLM-5.2 foundation, aimed at "research and engineering scenarios that require continuous environmental understanding, tool use, and multi-step task completion." The config the lab shipped tells the rest: 78 layers, 8 of 256 routed experts active per token, and a context window set to 1,048,576 tokens. Text input only — the tokenizer carries image, video and audio markers, but the chat template explicitly instructs the model to say it cannot accept media, so those are inherited plumbing rather than a multimodal claim.
What separates it from another big-weights drop is where the verification lives. The lab describes a Verifiable Experience Pipeline: every training task is connected to a real execution environment where the model observes state, calls tools, produces artifacts, and gets graded on external signals — passing tests, metrics, file state, source evidence. Only trajectories that tie a task, its steps, its output and its verification evidence together get absorbed as capability. That is the opposite of bolting agent skills on with prompt engineering at inference time, and it is the same bet the lab's chief scientist made publicly at the Pujiang forum — we covered his case for the science frontier in Shanghai AI Lab chief: science is the next programming.
The benchmark card compares Atria against DeepSeek V4 Pro, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 Sol and Claude Opus 5 across sixteen rows, and every number is vendor-reported. It claims the lead on five: BrowseComp 92.5 (against GPT 5.6 Sol's 92.2), CyberGym 86.5, DeepSearchQA 96.0, BFCL v4 77.0 and AutomationBench 53.8. It also shows its losses, which is what makes the table readable: SWE-bench Pro 59.6 against Opus 5's 74.7, JobBench 50.3 against 68.0, Terminal-Bench 2.1 78.3 against 90.2. The profile is a research-and-retrieval agent that is still mid-pack at engineering delivery.
The practical constraint is size. This is a weights release, not a service: running it today means downloading roughly 756 GB in FP8 or 1.5 TB in BF16 and serving it yourself, though the card ships deployment guides for SGLang and vLLM and hooks for Codex and Claude Code. An independent evaluation does not exist yet — Artificial Analysis has no page for the model — so treat the leaderboard rows as claims until someone neutral reproduces them.
What to watch: the first inference provider to serve a 744B MIT-licensed agentic model sets the real price of it, and the first neutral BrowseComp or CyberGym run decides whether the retrieval lead is real.
Is a silent weights drop better science communication than a launch event, or just a way to avoid questions about the training data? Tell us in the comments.
Sources: Atria Dawn Preview model card (Hugging Face) · Atria site · AILog — Shanghai AI Lab releases Atria Dawn Preview · OrcaRouter — Atria Dawn Preview ships quietly · 观点网 Guandian — Shanghai AI Lab open-sources ATRIA Dawn Preview, MetaX completes Day-0 adaptation · Atria-Dawn-Preview (GitHub)
