Shanghai Artificial Intelligence Laboratory has quietly released Atria Dawn Preview, a 744-billion-parameter mixture-of-experts agentic model, on Hugging Face under an MIT licence. The drop, dated the week of September 15, 2026, arrives without a splashy blog post or press event — the model card, an accompanying arXiv report and a 769-task human study are the entire launch package.
What Sits Inside
Atria Dawn is built on top of Shanghai AI Lab's own GLM-5.2 foundation model, with a 256K-token context window and FP8 quantised weights that make it practical to serve on a single 8-GPU node. The lab is deliberately positioning the release as an agentic model rather than a chat model: it is engineered for continuous environment interaction, tool use and multi-step task completion across four capability axes — discovery, creation, delivery and cybersecurity.
Benchmarks That Actually Matter For Agents
Atria's scorecard is heavy on the axes that separate a good chatbot from a working agent. It hits 96.0 on DeepSearchQA, 77.0 on BFCL v4 tool use, 53.8 on AutomationBench for long-horizon office and engineering workflows and 86.5 on CyberGym for security tasks. The 769-task human study behind the release covers real research, coding, document and security jobs — an evaluation methodology closer to what OpenAI and Anthropic have been publishing than a pure benchmark leaderboard.
Deployment
The Hugging Face repo ships SGLang and vLLM recipes and a direct Codex and Claude Code integration path. Shanghai AI Lab is exposing an API console for both international and China regions, mirroring the two-tier access model DeepSeek and Alibaba's Qwen team have used to keep the weights fully open while running a hosted service.
Open-Weights Continue To Close The Gap
The release lands amid a broader open-weights push out of China, where Unitree's blockbuster IPO, ByteDance spin-out Anew Labs' $290M raise, and Nutshell Therapeutics' AI drug discovery Series C1 have all pushed valuations for embodied and applied AI higher. It also lands the same week Shanghai AI Lab-adjacent research groups were named alongside NVIDIA and Salesforce for the Koa CRM reasoning model, suggesting Chinese labs are pivoting hard from foundation-model scale to agent-first design.
Reporting based on coverage from Pandaily, AI Weekly, Hugging Face and Shanghai AI Laboratory.