Shanghai AI Lab Unveils Atria Dawn: A Groundbreaking 744B Open-Source AI Agent

In an unexpected move that has taken the global AI community by surprise, the state-backed research institute Shanghai Artificial Intelligence Laboratory, renowned for the InternLM model family, has quietly unveiled Atria Dawn Preview on GitHub on September 11, 2026. This release was made without any press release, blog post, or pricing announcement. A subsequent FP8 checkpoint was released on Hugging Face a day later.

Atria Dawn Preview is a 744-billion-parameter mixture-of-experts (MoE) AI model. It is built on the robust GLM-5.2 foundation and has been released under the permissive MIT license. This means that developers have the freedom to download and self-host the open weights. The model boasts a 256K context window and is specifically designed for complex, long-horizon research and engineering tasks. These tasks include problem analysis, tool use, writing and running code, experiment execution, and failure recovery.

A comprehensive technical report, authored by over 140 researchers, was posted on arXiv on September 14. This report details the model’s unique Verifiable Experience Pipeline training approach, which grounds the model’s tool use in executable, checkable environments. According to the report, Atria Dawn Preview has achieved competitive results across 16 benchmarks, claiming top scores on five of them. Independent verification by third-party evaluators is currently underway.

Source: Startup Fortune – Shanghai AI Lab Quietly Releases Atria, a Free 744 Billion Parameter AI Agent