Shanghai AI Lab Releases Weights for Atria Dawn Preview Agentic Model

Shanghai AI Laboratory released a preview of the agentic model Atria Dawn Preview, based on GLM-5.2 (744B parameters, MoE, 256K context), which completes an open-ended task through the full cycle of 'analysis → plan → code → experiment → error analysis → retry,' embedding skills into the weights via a Verifiable Experience Pipeline.

Compiled by Sergey KostenchukPublished 2026-09-20Updated 2026-09-20

2026-09-20 Coding Anthropic

🛠 Shanghai AI Lab Releases Weights for Atria Dawn Preview Agentic Model

Atria Dawn Preview is out: the base is GLM-5.2 with 744B parameters (MoE, 256K context). The model solves a task through the full cycle of 'analysis → plan → code → experiment → error analysis → retry,' and skills are embedded directly into the weights via a Verifiable Experience Pipeline.

🌍 Skills—planning, tool use, and recovery from failures—are, for the first time, transferred from external scaffolds into the model's weights, and all of this is under the MIT license: leading on BFCL v4 (77.0), CyberGym—86.5, BrowseComp—92.5, but on SWE-bench Pro, Claude Opus 5 is ahead.

👤 You can try it right away: the API at api.atria-asi.ai connects to Codex, Claude Code, and Kimi Code; BF16 and FP8 weights are on Hugging Face and ModelScope; local deployment—SGLang and vLLM. Input is text-only—the API rejects images and PDFs.

Source 1: https://arxiv.org/abs/2609.15818 Source 2: https://github.com/atria-asi/Atria-Dawn-Preview

Sources