Atria Dawn Open Weights Land on Hugging Face

The Shanghai Artificial Intelligence Laboratory has published the full weights of Atria Dawn, its agentic model, on Hugging Face, with a 512,000-token context window and a README that says the code and weights are released under the MIT licence. The full release doubles the 256,000-token context of the preview weights the lab put out on 11 September.

The release closes a gap of about a week. When the lab announced the full model in mid-October it offered access only through its API, and the GitHub README still listed the full weights as “Coming soon” in its 19 October version. The Hugging Face repository went up early on 23 October, alongside an FP8 version, and the README on both sites now links to the downloads. For developers this turns a hosted model into one they can run, inspect and modify on their own hardware, which is the part of the story that is new today.

In brief

What’s new: Full Atria Dawn weights, plus an FP8 version, are public on Hugging Face, with a 512K context window and an MIT licence stated in the README. The lab’s claims: Its own table puts Atria Dawn first on six benchmarks, including CyberGym, but well behind Claude Opus 5 on SWE-bench Pro and NL2Repo. What’s missing: No independent benchmark runs yet, and no published hardware requirements.

The lab says it built Atria Dawn on GLM-5.2 for multi-step work

According to the README, Atria Dawn is a “new-generation agentic model developed by the Shanghai Artificial Intelligence Laboratory”, built on the “744B-parameter MoE GLM-5.2 foundation model”. Mixture-of-experts (MoE) models switch on only some of their parts for each token; the published configuration file lists 256 routed experts, eight of them used per token. The lab has not stated an active parameter count. It is the same institute behind the Intern-S2 science model that a recent maths preprint credited with helping develop a proof.

The files are large. Hugging Face lists 353 weight files totalling about 1.5 terabytes for the main release, and about 0.76 terabytes for the FP8 version. The lab points users to SGLang and vLLM for local deployment and gives setup guides for coding tools including Codex, Claude Code and Kimi Code. The model accepts text only; the README warns that image input is rejected with the message “Atria-Dawn is not a multimodal model”.

The lab pitches the model across four areas: discovery, creation, delivery and cybersecurity. It describes the cybersecurity work as “analyzing security issues, validating vulnerabilities, applying fixes, and performing re-validation in authorized environments”, a defensive framing that matters for how its strongest benchmark result should be read.

Its own table shows it leading on agent tasks and trailing on coding

All of the scores below come from the lab’s own evaluation table, which it says matches its website. No independent results have been published.

Benchmark	Atria Dawn	Claude Opus 5
AutomationBench	54.7	49.4
WideSearch	84.6	84.0
CyberGym	86.8	not listed
Terminal-Bench 2.1	81.0	90.2
SWE-bench Pro	61.7	76.5
NL2Repo-Bench	50.0	79.9

Source: Atria Dawn README, Shanghai AI Laboratory, 23 October 2026. All figures are the lab’s own.

On the lab’s numbers, Atria Dawn has the best score among the eight models compared on AutomationBench, CyberGym, SkillsBench (67.8), WideSearch, BFCL V4 (78.4) and the legal benchmark Harvey LAB (27.5). It is behind on software engineering. Claude Opus 5 scores 76.5 on SWE-bench Pro to Atria Dawn’s 61.7, and 79.9 on NL2Repo-Bench to its 50.0. Several Chinese rivals also edge it on coding in the same table: Qwen 3.8 Max scores 65.3 on SWE-bench Pro and Kimi K3 62.0.

The CyberGym figure, 86.8, compares with 84.5 for GLM 5.3 and 83.6 for GPT 5.6 sol in the table, while no Claude Opus 5 score is given. The full release improves only slightly on the preview in the lab’s own testing: 86.5 to 86.8 on CyberGym and 59.6 to 61.7 on SWE-bench Pro.

A permissive licence matters more than any single benchmark row

The most consequential line in the README may be the licence, not the scores. The MIT licence lets anyone use, modify and redistribute the weights, including commercially, with little more than an attribution notice. That sets Atria Dawn apart from releases such as MiniMax’s M3.1 Flash, whose weights are on Hugging Face under a non-commercial community licence. It also puts a model of this size in the same permissive bracket as much smaller releases like Recursive’s Apache-licensed Flint. One housekeeping point: the MIT text itself sits in the GitHub repository, while the Hugging Face listing relies on the README’s statement and carries no licence tag.

The cyber score deserves careful reading rather than alarm. Epoch AI estimated last week that open-weight models trail closed ones on cyber skills by six to seven months, with GLM-5.3 the top open model on its index. If the lab’s CyberGym result holds up, defenders gain a capable vulnerability-analysis model they can run inside their own networks, without sending sensitive code to an outside API. The same openness means the capability is available to anyone, and unlike a hosted model, downloaded weights cannot be patched or withdrawn later. That is why independent evaluation of open releases like this one matters: the lab’s table is a starting point, not a verdict, and it gives no CyberGym score for Claude Opus 5.

Outside tests of the full weights are the next check

With the weights public, researchers and evaluation groups can now run Atria Dawn themselves rather than relying on the lab’s figures or its API. The first useful results will be independent runs of SWE-bench Pro and CyberGym on the full 512K release, and a clear statement from the lab of the hardware needed to serve it.