The agentic model appeared on Hugging Face three days before the scientific article, without a blog post or press release, under an MIT license and with a free API in preview.
The Shanghai Artificial Intelligence Lab has released Atria Dawn Preview, a preview of an agentic model built on the GLM-5.2 expert-mix base and its 744 billion parameters. The release was quiet: the Hugging Face repository appeared on September 11, followed by an FP8 checkpoint the next day, with the scientific paper arriving only three days later — without a blog post or press release.
The model is aimed at research and engineering uses that involve a continuous understanding of the environment, the use of tools and the sequence of tasks over several steps. The claimed loop covers the analysis of the problem, the design of the solution, the call of tools, the implementation of the code, the execution of the experiment, the analysis of the results and the recovery after failure, by constantly confronting the objective with the return of the environment.
Training involves what the authors call a verifiable experience pipeline, which anchors the use of the tools in truly executable environments. The article announces competitive results on sixteen evaluation sets, including five best scores, and is based on 769 task reports produced by 56 participants; about a third of the assisted tasks were deemed unfeasible without the agent. The model is released under an MIT license, accompanied by a free OpenAI-compatible API in preview.
Read the full article on the original source
Go to the source →