Atria Dawn Preview is a 744 billion parameter mixture-of-experts model that Shanghai Artificial Intelligence Laboratory released as an open-weight checkpoint under an MIT license, with roughly 40 billion parameters active per token. It is built on Zhipu AI's GLM-5.2 foundation, which was originally shipped in June 2026, and then post-trained by Shanghai AI Lab specifically for long-running agentic work: problem analysis, tool use, writing and executing code, running experiments, and recovering from mid-step failures. Both BF16 and FP8 weight files are available for self-hosting, and the same model is reachable through a hosted API that mirrors common Chat Completions, Messages, and Responses interfaces, so existing agent stacks can be pointed at it by swapping an endpoint rather than rewriting the integration.
Benchmark coverage focuses on agentic and coding tasks, with thirteen source-displayable rows currently published and several categories left unmeasured. The model sets the best verified result on BrowseComp, DeepSearchQA, and MLE-Bench Lite among tracked peers, while posting strong tool-use numbers on CyberGym and BFCL v4 and a more middling 53.8 on AutomationBench. On coding workloads it trails the frontier, landing behind top verified peers on SWE-bench Pro and Terminal-Bench 2.1, which makes it a better fit for agentic research, browser-driven investigation, and computer-use pipelines than for teams that need a pure coding leader. The benchmark figures are vendor-published and have not yet been independently reproduced, and the released artifact is explicitly a preview with no announced follow-up release date.
