Shanghai AI Lab Ships Atria Dawn Full Release With a 512K Context Window

Shanghai Artificial Intelligence Laboratory released the full version of Atria Dawn in mid-October, roughly a month after the Atria Dawn Preview appeared in mid-September. The model remains positioned as an agent-oriented general-purpose system for research and engineering work that requires reading an environment, calling tools and carrying a task through multiple steps.

The headline change is context length, which rises from 256K tokens in the preview to 512K in the full release. The project's GitHub repository complicates that number slightly: its model download table still lists 256K for both the full model and the FP8 build, so developers should check the hosted API documentation for the limit that actually applies.

The base is unchanged. Atria Dawn is built on GLM-5.2, a 744-billion-parameter mixture-of-experts model, and the lab now describes its abilities in four buckets: discovery, creation, delivery and cybersecurity. Discovery covers evidence gathering, deep research and turning research questions into executable experiment plans. Creation spans software, interactive applications, games, data visualizations and machine-learning systems. Delivery converts documents, data and design requirements into reports and slide decks. Cybersecurity work is limited to authorized environments and covers analysis, verification, remediation and retesting.

Against the preview, the lab highlights three gains: JobBench rises to 57.5 from 50.3, a 7.2-point improvement; WideSearch reaches 84.6 from 81.9; and PaperBench climbs to 89.6 from 87.0.

The broader benchmark table is more mixed. The repository's evaluation table, which the lab says matches its website, lists AutomationBench at 54.7, SkillsBench at 67.8 and SWE-bench Pro at 61.7, with smaller improvements on most other tests. On BrowseComp the preview actually leads, 92.5 to 91.3. The pattern suggests a targeted step forward rather than a uniform capability jump.

Access is split between hosted and self-hosted paths. Atria Dawn is available through APIs with separate international and China endpoints, and the repository ships setup guides for SGLang, vLLM and Codex. It takes text input only, and the code is released under the MIT license.

Weights are the open question. The full release is still marked "Coming soon" on Hugging Face and ModelScope; only preview weights can be downloaded today. For now, the open-source label applies to code, deployment scaffolding and architecture documentation rather than to parameters a team can fine-tune.

A one-month gap between preview and full release suggests the lab is treating agent models as a flagship line rather than an experiment. The things to watch are concrete: when the full weights land, and whether 512K context holds up in long-document research, repository-scale code understanding and multi-step tool use outside the benchmark harness.