Shanghai AI Lab Quietly Opens Atria Dawn Preview, a 744B MIT Agentic MoE

Shanghai AI Laboratory’s InternLM organization quietly published Atria Dawn Preview on Hugging Face in mid-September 2026, shipping open weights under the MIT License for a large agentic mixture-of-experts instruct model. The Hugging Face card for internlm/Atria-Dawn-Preview describes a preview built on a 744-billion-parameter MoE foundation aimed at multi-step tool use, experiment execution, and recoverable agent workflows, with a companion FP8 checkpoint at internlm/Atria-Dawn-Preview-FP8.

The model card’s download table lists a 256K context window for both BF16 and FP8 instruct builds. The README frames four delivery-oriented scopes—Discovery, Creation, Delivery, and Cybersecurity—and documents local serving paths through SGLang and vLLM plus optional hosted access notes. As of a September 14 secondary write-up, the lab had not published a separate blog announcement or technical paper alongside the Hub drop.

Vendor benchmarks and caveats

The card includes a vendor-reported comparison table against DeepSeek V4 Pro, Kimi K3, Qwen3.8-Max, GLM-5.3, GPT-5.6 Sol, and Claude Opus 5. Highlight rows on that table include BrowseComp 92.5, CyberGym 86.5, and DeepSearchQA 96.0, with weaker coding-agent rows such as Terminal-Bench 2.1 at 78.3 and SWE-bench Pro at 59.6. Treat those figures as lab-reported until independent evaluations appear.

This brief is about the quiet Atria Dawn Preview weights drop, not the same lab’s Apache-2.0 Intern-S2-397B scientific multimodal release.

Primary sources are the Hugging Face model cards for Atria-Dawn-Preview and Atria-Dawn-Preview-FP8, with September 14 secondary context from OrcaRouter’s quiet-release report.