Shanghai AI Lab Launches Atria Dawn Full Release with 512K Context
Shanghai AI Lab rolled out the full version of Atria Dawn with a 512K API context and benchmark gains, while standalone weights remain pending.
Image source: https://github.com/atria-asi/Atria-Dawn/blob/main/assets/evaluation.jpg
Full Release Delivers Doubled Context Length
Shanghai AI Lab has deployed the full release of Atria Dawn, an agentic model structured for multi-stage research and software engineering pipelines. Built on the 744-billion-parameter GLM-5.2 Mixture of Experts foundation, the update targets end-to-end task execution from problem analysis through verification.
The official API expands context capacity to 512K tokens, doubling the 256K ceiling of the earlier preview edition. This window includes both input context and generated output, where output tokens can range up to 65,536 depending on stop conditions. Operational tasks can now draw on longer reference histories, though teams must balance system instructions and tool responses against available generation space.
Benchmark Results Show Widespread Improvements
Official evaluation records from Shanghai AI Lab show measured gains across 17 of 20 benchmark metrics compared directly against preview scores. The largest reported gain arrived on JobBench, where real-world operational evaluation rose from 50.3 to 57.5.
Other areas showing higher marks include PaperBench, WideSearch, and SWE-bench Pro. However, performance did not improve across every metric. The Harbor-Index score remained flat at 17.3, while BrowseComp and NL2Repo-Bench recorded slight decreases, showing specific trade-offs across workflows.
Weights Remain Pending for Local Deployment
Despite availability through the hosted endpoint, full-version model weights have not been released for local or self-hosted infrastructure. The project repository marks full and FP8 weights as coming soon, leaving direct execution restricted to the earlier preview download.
In addition, the official distribution table lists 256K for upcoming self-hosted weights, diverging from the 512K supported on the hosted service. Until local weights arrive, access remains limited to the provider's API.