The Pulse
Shanghai AI Lab Releases Atria Dawn, a 744B-Parameter Agent Model
Shanghai AI Lab’s Atria Dawn Preview, built on the 744-billion-parameter GLM-5.2, is an open-weight agentic model for research and engineering. Its paper reports top scores on five of 16 benchmarks and studies how people and agents shared t
AI.info Team · Written with AI · fact-checked ·
Shanghai Artificial Intelligence Laboratory is pitching Atria Dawn Preview as a model for work that takes more than a prompt and a reply. The agentic model is built on Z.ai’s GLM-5.2, a 744-billion-parameter mixture-of-experts foundation model, and targets research and engineering tasks that involve using tools, running code, carrying out experiments and recovering when a step fails. The lab’s Hugging Face card offers the weights under the MIT License.
From tool calls to checked results
The model is designed for extended workflows: analyzing a problem, planning a solution, calling tools, implementing code, executing experiments and reviewing results. The card groups its intended uses into discovery, creation, delivery and cybersecurity, with security work limited to authorized environments. It lists a 256K-token context window and a separate FP8-quantized version.
The accompanying paper, posted to arXiv on September 14, describes the training approach as a Verifiable Experience Pipeline. Tasks connect the model to execution environments where it can observe state, use tools and produce intermediate artifacts; the final result is checked against external signals such as tests, metrics, file state or source evidence. Recurring failures, such as poor tool selection or incomplete verification, are turned into new tasks and quality checks.
Five top scores in the lab’s 16-benchmark table
The lab reports the highest score among the compared models on five of 16 benchmarks: 96.0 on DeepSearchQA, 92.5 on BrowseComp, 77.0 on BFCL v4, 53.8 on AutomationBench and 86.5 on CyberGym. The comparison set is six models: DeepSeek V4 Pro, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT-5.6 sol and Claude Opus 5, and the paper notes that the figures are those published on the official release website.
The same table shows weaker results on several software and workplace tests. Atria scores 59.6 on SWE-bench Pro, below the listed 74.7 for Claude Opus 5, and 78.3 on Terminal-Bench 2.1, below Qwen 3.8 Max’s 89.3 and Claude Opus 5’s 90.2. On Workspace-Bench, Atria’s 65.0 trails Claude Opus 5’s 65.8. The strengths cluster in search, tool calling and security rather than forming an across-the-board lead.
The paper studies how people and agents built Atria
Much of the report is about the development project itself. The team analyzed 769 task records and agent logs from 56 participants; of the 739 tasks with a clear answer on AI use, 713, or 96.5%, involved AI. Among 455 completed AI-assisted tasks with a usable response, participants rated 151, or 33.2%, as infeasible without AI under the same scope and resources.
The records point to a division of labor rather than autonomy. Of 588 tasks that hit a difficulty, 76.0% moved forward because a person intervened. The most common forms of help were adding context or clarifying requirements (35.2%) and diagnosing the issue or changing the method (34.7%); partial edits by a person made up 3.2% and full takeovers 0.7%. From that record, the paper’s project leads, Honglin Guo and Tao Gui, and their co-authors drew a conclusion about oversight. They wrote in the abstract:
“Progress toward more autonomous AI research must therefore advance both the capacity for discovery and the capacity for meaningful human oversight, preserving accountable human authority over the risks and direction of continued development.”
Honglin Guo and Tao Gui, project leads, and co-authors, in the Atria Dawn paper
The paper lists 143 authors, credited as the Atria Team, with both correspondence addresses at Fudan University. The weights for the standard and FP8 versions are available now on Hugging Face under MIT terms.