Pandaily reported on 17 October that Shanghai Artificial Intelligence Laboratory had released the full version of Atria Dawn, an agent model built on a 744-billion-parameter mixture-of-experts GLM-5.2 base, about a month after its September preview.
The laboratory says the model’s context window has grown from 256K to 512K tokens. That window is the text a model can consider at once, much like the pages spread across a desk while working through a problem. The project’s GitHub download table, however, lists 256K for both the full model and its FP8 version.
Working through a long research document could require splitting it into fewer portions if the larger limit applies.
The laboratory reports scores of 57.5 on JobBench, 84.6 on WideSearch and 89.6 on PaperBench, against 50.3, 81.9 and 87.0 respectively for the preview. Its evaluation table also gives the full version 91.3 on BrowseComp, below the preview’s 92.5.
The model accepts text input through hosted APIs, and its code is released under the MIT licence. Preview weights can be downloaded, but full-version weights are listed as ‘Coming soon’ on Hugging Face and ModelScope.
