The Pulse
Shanghai AI Lab Gives Atria Dawn a 512K API Context
Shanghai AI Laboratory has released the full version of Atria Dawn, an agent model whose API documentation lists a 512K-token context window. Its own benchmark table shows higher scores than the Preview version on 17 of 20 comparable tests,
AI.info Team · Written with AI ·
Atria Dawn’s full release scores higher than its Preview version on 17 of 20 comparable benchmarks, according to Shanghai Artificial Intelligence Laboratory’s evaluation table. The update also brings a 512K-token context window to the model’s hosted API—but a mismatch in the lab’s own documentation leaves the context limit for downloadable weights unclear.
A 512K API, but a 256K listing
The lab describes Atria Dawn as an agent model for research and engineering tasks that involve following a problem through several steps: gathering evidence, using tools, writing code, running experiments and checking results. It is built on GLM-5.2, a 744-billion-parameter mixture-of-experts model, according to the project repository.
The API documentation lists Atria-Dawn with a 512K-token context window. That capacity covers the entire request, including instructions, conversation history, documents and tool results, as well as generated text. The documentation sets the output-token limit at no more than 65,536, so 512K is not the amount the model can generate in a single response.
The repository’s download table, however, lists 256K for both the full model and its FP8 version. Both entries are marked “Coming soon” on Hugging Face and ModelScope; the Preview weights are the versions with download links. The public materials do not explain why the API’s stated limit differs from the repository’s figure, or confirm that a self-hosted full release will support 512K.
JobBench rises 7.2 points
The lab’s table records the largest improvement on JobBench, where Atria Dawn scores 57.5 against 50.3 for Preview, a gain of 7.2 points. It also reports gains on WideSearch, from 81.9 to 84.6, and PaperBench, from 87.0 to 89.6. The figures are the lab’s own evaluations, not independent verification.
The results are not uniformly higher. Atria Dawn’s BrowseComp score falls from 92.5 to 91.3, and its NL2Repo-Bench score drops from 51.4 to 50.0. Harbor-Index stays at 17.3. The 17-of-20 tally counts benchmark entries with results for both versions; three entries without a Preview result are not part of that comparison.
Nor does a higher score than Preview make Atria Dawn the top scorer in every comparison shown. On JobBench, the same table lists GLM 5.3 at 58.2 and Claude Opus 5 at 68.0, above Atria Dawn’s 57.5. Those comparisons come from the lab’s published table, and its figures should be read as the developer’s reported results.
Built for work that spans several steps
The Shanghai AI Laboratory says Atria Dawn targets four areas: Discovery, Creation, Delivery and Cybersecurity. The descriptions range from research and evidence gathering to software development, reports and presentations, and security work in authorized settings. The model accepts text input; the repository warns that image inputs are not supported.
“The model is designed to drive open-ended problems toward executable, verifiable, and reproducible results,” the Shanghai Artificial Intelligence Laboratory writes in its Atria Dawn repository. The lab’s stated goal is to connect tool use and environmental feedback across a task, rather than treat each prompt as a standalone answer.
Hosted access comes before local weights
The API documentation provides Chat Completions, Messages and Responses interfaces, with the same model ID, Atria-Dawn, for each. That gives developers a way to connect existing applications to the hosted service. But the full model’s listed weights are still unavailable, leaving the downloadable Preview as the only version with weights linked from the repository.
The immediate distinction is practical: developers can call the full release through the API, whose documentation states a 512K context, but cannot yet check that capability against the full model running on their own hardware. Until the full weights are posted and the 256K-versus-512K discrepancy is clarified, the repository does not establish whether local deployments will match the hosted API’s context limit.