Shanghai AI Laboratory · 2026-09-14 · major
Atria Dawn Preview — the report behind Shanghai AI Lab's 744B agent model
Shanghai AI Laboratory published the technical report for Atria Dawn Preview, the 744B MoE agentic model it released quietly under MIT on 11 September. More than 140 authors describe the training pipeline and 16 benchmark results.
The 744B open-weight agent model that showed up on GitHub with no announcement now has a 140-author paper explaining how it was built.
Quick facts
| Maker | Shanghai AI Laboratory (InternLM) |
|---|---|
| Size | 744B MoE, built on GLM-5.2 |
| Context window | 256K tokens |
| License | MIT |
| Weights | Hugging Face + ModelScope, BF16 and FP8 |
| Serving | SGLang v0.5.13.post1+ or vLLM v0.23.0+ |
| What's new | Technical report, arXiv 14 September 2026 |
Benchmarks
| GPT-5.6 | 88.9% | |
|---|---|---|
| Claude Opus 5 | 88% | |
| DeepSeek V4 | 86.8% | |
| Atria Dawn Preview | 86.2% | |
| Kimi K3 | 85.8% | |
| Qwen 3.8 | 81.3% | |
| GLM 5.3 | 80.8% |
What is it?
Atria Dawn Preview finally has a technical report. The model itself appeared on 11 September 2026 with no blog post and no paper — just MIT-licensed weights on Hugging Face and ModelScope from Shanghai AI Laboratory, the public research lab behind the InternLM family. The report, posted to arXiv on 14 September with more than 140 authors, is the first full account of what the model is and how it was trained.
How does it work?
Training used what the authors call a Verifiable Experience Pipeline, which connects tool-mediated interactions to executable environments so the model learns from actions that can be checked rather than from text alone. The base is the 744B-parameter GLM-5.2 mixture-of-experts model with a 256K context window. The report groups results into four areas — Discovery, Creation, Delivery and Cybersecurity — across 16 benchmarks, and adds a study of 769 human-AI task records.
Why does it matter?
An open-weight agent model at this scale is rare, and MIT terms mean you can run and modify Atria Dawn Preview without asking anyone. The report closes the gap that made the quiet release hard to judge: the scores are now traceable to a method, and the human-AI study describes a working pattern where the agent proposes methods and implements revisions while people make the final calls.
Who is it for?
ML engineers and researchers running self-hosted agents
Frequently asked questions
- Is Atria Dawn Preview free to use commercially?
- Yes. Shanghai AI Laboratory publishes Atria Dawn Preview under the MIT License, which permits commercial use, modification and redistribution. Both the full-precision and FP8 checkpoints are available on Hugging Face and ModelScope, so you can self-host without a licence negotiation or usage reporting.
- How does Atria Dawn Preview compare to Kimi K3 and GPT-5.6?
- Atria Dawn Preview leads on two of the three benchmarks with full comparison tables: 96.0 on DeepSearchQA against 95.9 for Kimi K3 and 93.2 for GPT-5.6, and 77.0 on BFCL v4 tool use against 74.1 for GLM 5.3. On MLE-bench Lite it scores 86.2, behind GPT-5.6 at 88.9 and Claude Opus 5 at 88.0.
- What do I need to run Atria Dawn Preview myself?
- Atria Dawn Preview is a 744B mixture-of-experts model, so serving it needs a multi-GPU setup. Shanghai AI Laboratory supports two inference stacks: SGLang from v0.5.13.post1 and vLLM from v0.23.0, both with published recipes for the GLM-5.2 base it is built on. An FP8 checkpoint is offered to cut memory needs.
- Can I call Atria Dawn Preview through an API instead of self-hosting?
- Yes. Shanghai AI Laboratory runs hosted endpoints for Atria Dawn Preview at api.atria-asi.ai/v1 for international traffic and discovery.intern-ai.org.cn for China, using the model id Atria-Dawn-Preview. The report and repo document the endpoints rather than a price list, so self-hosting is the option with known costs.
Try it
Weights: huggingface.co/internlm/Atria-Dawn-Preview — serve with vLLM v0.23.0+ or SGLang v0.5.13.post1+