New model · Shanghai Artificial Intelligence Laboratory
Atria Dawn Preview
744B-parameter MoE agentic model trained on verified tool-use trajectories, tops 5 of 16 benchmarks
· 12 claims checked against 2 sources
What it is
Atria Dawn Preview is a foundation agentic language model released by Shanghai Artificial Intelligence Laboratory. It is built on GLM 5.2, a 744 billion parameter mixture-of-experts base model, and targets scientific research and engineering workflows that require sustained tool use, multi-step planning, and recovery from failed attempts.
How it works
The model is trained through what the team calls a Verifiable Experience Pipeline. Every training task runs in a real environment, and outcomes are checked against tests, metrics, file state, geometric structure, or source evidence, rather than only human preference, before that trajectory is kept as training data.
What is new
- Highest reported score on 5 of 16 benchmarks, including DeepSearchQA (96.0), BrowseComp (92.5), AutomationBench (53.8), and BFCL v4 (77.0)
- On CyberGym it scores 86.5, beating the runner-up GLM 5.3 (84.5) by 2.0 points, indicating particular strength in vulnerability analysis
- A 256K context window, released as both a full precision and an FP8-quantized checkpoint
- A companion study of the model's own development draws on 769 task records from 56 participants, who rated roughly one third of completed AI-assisted tasks as infeasible without AI
Limits
Atria Dawn Preview is not uniformly ahead of other models. On SWE-bench Pro, Terminal-Bench 2.1, and JobBench, it trails Claude Opus 5 by meaningful margins, indicating that its strengths in search and tool use do not extend evenly across every agentic task.
Availability
Weights and code are released under the MIT License on Hugging Face and ModelScope. Shanghai Artificial Intelligence Laboratory also provides a hosted API and a paper describing the model on arXiv.
The evidence
Each statement above was written with the sentence from the source that supports it, and a program (not a model) checked that every quote and every figure appears in the pages it read. Here they are.
- 01
Atria Dawn Preview is built on a 744B-parameter MoE GLM-5.2 foundation model
“Built on the **744B-parameter MoE GLM-5.2** foundation model, it targets research and engineering scenarios” S1
- 02
The model supports a 256K context window and is available in full precision and FP8-quantized versions
“Atria-Dawn-Preview-FP8 | FP8-quantized Instruct model | 256K” S1
- 03
Atria Dawn Preview achieves the highest reported score on 5 of 16 benchmarks
“Atria Dawn achieves the highest reported score on five benchmarks and the second-highest on three others” S2
- 04
It scores 96.0 on DeepSearchQA and 92.5 on BrowseComp
“DeepSearchQA (Gupta et al., 2026) (96.0), and BrowseComp (Wei et al., 2025) (92.5)” S2
- 05
It scores 53.8 on AutomationBench and 77.0 on BFCL v4, the highest reported
“AutomationBench (Shepard and Salimans, 2026) (53.8), BFCL v4 (Patil et al., 2025) (77.0)” S2
- 06
On CyberGym it scores 86.5, exceeding the runner-up GLM 5.3 by 2.0 points
“it obtains the highest reported score on CyberGym at 86.5, exceeding the runner-up GLM 5.3 (84.5) by 2.0 points” S2
- 07
Atria Dawn is trained via a Verifiable Experience Pipeline connecting tool-mediated interactions to executable environments and externally verified outcomes
“This model is trained via a Verifiable Experience Pipeline that connects tool-mediated interactions to executable environments and externally verified outcomes” S2
- 08
A study of the model's own development draws on 769 task records from 56 participants alongside agent logs
“Drawing on 769 task records from 56 participants alongside agent logs, we examine how responsibilities were distributed between humans and agents across the project” S2
- 09
Participants rated roughly one-third of completed AI-assisted tasks as infeasible without AI
“Participants rated roughly one-third of completed AI-assisted tasks as infeasible without AI under the same scope and resource constraints” S2
- 10
On SWE-bench Pro, Terminal-Bench 2.1, and JobBench, Claude Opus 5 scores higher than Atria Dawn Preview
“SWE-bench Pro | 59.6 | 58.3 | 61.6 | 65.1 | 60.3 | 61.4 | **74.7**” S1
- 11
The code and model weights are released under the MIT License
“The code and model weights in this repository are released under the [MIT License](LICENSE)” S1
- 12
Atria Dawn Preview is available via hosted API access as well as local deployment
“Atria Dawn Preview supports both local deployment and hosted access.” S1
