Shanghai AI Lab Releases Atria Dawn With 512K-Token Context

The Shanghai Artificial Intelligence Laboratory has released the full version of Atria Dawn, and its headline number is now 512K tokens, the amount of text the model can keep in view at once. The preview offered 256K. Developers who open the project's GitHub repository will still find 256K in the download table, for both the full model and its FP8 version, a compressed 8-bit variant. The hosted API documentation is the place to check which limit applies.

Atria Dawn is an agent-oriented model: it is meant to read an environment, call tools and carry a task through several steps rather than answer a single question. It is trained on GLM-5.2, a base model of 744 billion parameters organized as a mixture of experts, a design in which the network is split into specialist sub-networks and only some of them respond to each request. The lab sorts its abilities into four families: discovery (gathering evidence, deep research, turning a question into an executable experiment plan), creation (software, games, data visualizations, machine-learning systems), delivery (reports, slide decks, structured documents) and cybersecurity, which covers finding and retesting vulnerabilities only in authorized environments. The full model arrived about a month after the preview, in mid-September.

The lab's own numbers show the largest gains on JobBench, 57.5 against 50.3 for the preview, followed by WideSearch (84.6 against 81.9) and PaperBench (89.6 against 87.0). On most other benchmarks, including AutomationBench (54.7), SkillsBench (67.8) and the software-engineering test SWE-bench Pro (61.7), the changes are smaller. A few results went the other way: the preview scored 92.5 on BrowseComp against 91.3 for the full version. These figures come from the laboratory, which says they match its website; no independent evaluation is cited in the sources, and they should be read as the lab's own measurements.

For developers and research teams, the change is immediate on one front and pending on another. Hosted APIs are open now, with separate international and China endpoints, and the repository includes setup guides for SGLang, vLLM and Codex. The code is released under the MIT license. But the full model's weights are still listed as 'Coming soon' on Hugging Face and ModelScope, so anyone who wants to run the complete model on their own hardware has to wait; only the preview weights can be downloaded today. The model accepts text input only.