Shanghai AI Lab opens Atria Dawn Preview for long-horizon research agents

ATRIA, from the Shanghai Artificial Intelligence Laboratory, released Atria Dawn Preview — an open agentic model built to carry scientific work from a method in the literature to executable experiments, reproducible metrics, and an inspectable report. The model card and weights are live on Hugging Face; documentation and evaluations are published at atria-asi.ai.

SOFTWARE desk — same-day primary open release of a large Chinese lab agent model aimed at end-to-end research automation, with public weights and reported tool-use/security benchmarks.

Built on a 744-billion-parameter mixture-of-experts foundation the lab names GLM-5.2. Mixture-of-experts, or MoE, means a large model that turns on parts of itself for each query instead of running every parameter every time. The lab says the model is built for continuous tool use, multi-step tasks, and results other people can check.

The same model card positions the work across four workflows: Discovery — turning a research question into an executable experiment plan. Creation — building software, apps, and machine-learning systems. Delivery — turning documents and data into reports and other structured files. Cybersecurity — analyzing security issues and checking fixes in authorized environments.

Lab-reported scores on the Hugging Face model card — not independent Bad Signal runs: AutomationBench 53.8; CyberGym 86.5; BFCL v4 77.0; DeepSearchQA 96.0; BrowseComp 92.5. AutomationBench is a tool-use automation bench. CyberGym is a security-task bench. BFCL is the Berkeley Function-Calling Leaderboard, a tool-calling test. The same card and the Newsfile note that some individual metrics may change as evaluation continues, and that the project site carries current comparison tables.

Availability, lab: instruct and FP8 weights are on Hugging Face at launch under an MIT license. MIT here means you can use, change, and share the weights, including commercially, if you keep the license notice. The lab also publishes local-deploy notes for SGLang and vLLM and hosted API endpoints by region. Version pins and install commands stay in Sources.

Plain English for the rest of the card: agentic = the model is meant to use tools and keep working across many steps, not just answer one chat turn. MoE = mixture-of-experts, a big model that activates parts of itself per query. open weights = downloadable model parameters you can run yourself. MIT = a permissive open-source license. Preview here is the product name, not a closed beta.