Shanghai Artificial Intelligence Laboratory has released the full version of Atria Dawn, its agent-oriented large model, about a month after the Atria Dawn Preview appeared in mid-September. The model is trained on the 744-billion-parameter mixture-of-experts GLM-5.2 base and targets research and engineering work that requires understanding an environment, calling tools and completing multi-step tasks.
The headline change in the lab's announcement is the context window, which grows to 512K tokens from the 256K offered by the preview. The model download table in the project's GitHub repository still lists 256K for both the full model and its FP8 version, so developers may want to check the hosted API documentation for the limit that applies.
The lab now frames the model's abilities in four categories. Discovery covers gathering and organizing evidence, deep research and turning research questions into executable experiment plans. Creation covers building software, interactive applications, games, data visualizations and machine-learning systems. Delivery turns documents, data and design requirements into reports, slide decks and other structured outputs. Cybersecurity covers analyzing security issues, verifying vulnerabilities, applying fixes and retesting, limited to authorized environments.
Compared with the preview, the lab highlights gains of 7.2 points on JobBench (57.5 versus 50.3), 2.7 points on WideSearch (84.6 versus 81.9) and 2.6 points on PaperBench (89.6 versus 87.0). The repository's evaluation table, which the lab says matches its website, shows smaller improvements on most other benchmarks, including AutomationBench at 54.7, SkillsBench at 67.8 and SWE-bench Pro at 61.7. On a few tests the preview scored equal or higher, such as BrowseComp, where the preview's 92.5 tops the full version's 91.3.
The model is available through hosted APIs with separate international and China endpoints, and the repository includes setup guides for SGLang, vLLM and Codex. It accepts text input only. The code is released under the MIT license, but weights for the full release are still listed as 'Coming soon' on Hugging Face and ModelScope; only the preview weights can be downloaded today.