Shanghai AI Lab Releases Atria Dawn Preview, an Open-Source Research Agent Foundation Model
On September 14, Shanghai AI Laboratory, together with Fudan University, the Institute of Software of the Chinese Academy of Sciences, Renmin University and other institutions, released Atria Dawn Preview, a foundation model for scientific research and complex agent tasks, simultaneously opening its website, API, GitHub, Hugging Face and ModelScope, with model weights and technical report fully open-sourced.
Confirmed
- The model is built on a 744B-parameter MoE GLM-5.2 base (per blogger lxfater) and focuses on verifiable delivery in long-horizon tasks: blogger xiaohu relayed that its training relies on a "verifiable experience pipeline," where only experience tested in real environments, validated by metrics, and consolidated is used for training.
- It is positioned as a research productivity model, aimed at continuously handling complex research tasks and turning long-cycle work into step-by-step verifiable outcomes rather than solving single-point tasks (blogger jiqizhixin).
- The model flags unverifiable information in its answers; blogger aliscodes tested this by asking it to find 2024 sales data, confirming the feature.
- It took first place in two of eight benchmark comparisons (AutomationBench, CyberGym) and second in two more (per blogger 新智元); 新智元 also noted that 96.5% of R&D tasks were completed by AI.
- On the team front, 65 of the 96 core members are current students (bloggers xiaohu, 新智元).
Why it matters
- The model attempts to mitigate hallucination and unreliability in long-horizon agent tasks through "verifiable experience" and proactive uncertainty labeling, a significant practice in the research-agent direction.
- The large team with two-thirds students and the fully open-source strategy (weights + technical report + API) are rare among domestic LLM releases, making its future iterations worth watching.
2026-09-14 ~ 2026-09-15 · 7 related posts
Primary sources
- Shanghai AI Lab's Atria-Dawn-Preview agent ranks top on agent benchmarks while AI ran 96.5% of its own training — 新智元 ·
- Shanghai AI Lab open-sources Atria Dawn, a 744B MoE agentic model built for verifiable delivery — lxfater ·
- Shanghai AI Lab's Atria Dawn: a research agent that flags what it can't verify — aliscodes ·
- [source] Shanghai AI Lab's Atria-Dawn-Preview agent ranks top on agent benchmarks while AI ran 96.5% of its own training — 新智元 · 2026-09-14
- [source] Shanghai AI Lab's Atria Dawn: a research agent that flags what it can't verify — aliscodes · 2026-09-14
- Shanghai AI Lab and universities release Atria Dawn Preview for long-horizon scientific work — jiqizhixin · 2026-09-15
- Atria Dawn Preview: research-focused base model trained on a verified-experience pipeline — xiaohu · 2026-09-15
- Student-heavy AtriaASI debuts Atria Dawn, a base model trained on verified real-world experience — xiaohu · 2026-09-15
- [source] Shanghai AI Lab open-sources Atria Dawn, a 744B MoE agentic model built for verifiable delivery — lxfater · 2026-09-15
1 near-duplicate retellings: xiaohu