Atria Dawn
Shanghai AI Laboratory released the full version of Atria Dawn, a 744B MoE agent model on a GLM-5.2 base, with the context window raised to 512K tokens from 256K.
The full version follows the September preview and adds a 512K context window, though the GitHub download table still lists 256K. It scores 57.5 on JobBench against 50.3 for the preview, 84.6 on WideSearch against 81.9, and 89.6 on PaperBench against 87.0. On a few tests the preview scored higher, such as BrowseComp (92.5 versus 91.3). It is available through hosted APIs, and the full weights are still listed as coming soon.
- Date
- Saturday 17 October 2026
- Lab
- Shanghai AI Laboratory (InternLM)
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| JobBench | 57.5 preview scored 50.3 | company |
| WideSearch | 84.6 preview scored 81.9 | company |
| PaperBench | 89.6 preview scored 87.0 | company |
| SWE-bench Pro | 61.7 | company |
| BrowseComp | 91.3 preview scored 92.5 | company |
Code is MIT licensed and the preview weights are open, but full-release weights are listed as 'Coming soon' on Hugging Face and ModelScope. Text input only.
Sources
This record was checked against its sources on 18 October 2026. How we check