Atria Dawn is an open agentic model from the Shanghai Artificial Intelligence Laboratory, built for research and engineering jobs that take many steps and tools.
In plain words
Atria Dawn is built on GLM-5.2, a mixture-of-experts model with 744 billion parameters, and is tuned to keep going through long jobs: watching its surroundings, using tools and finishing step after step. It was trained on experience whose outcomes were checked by running code and other outside tests, and it aims for results others can repeat. It reads text only, up to 512K tokens. The weights use the MIT licence, and you can run it yourself or use a hosted API.
A quick example
Imagine a researcher asks Atria Dawn to rerun an experiment from a paper. The model reads the paper, writes and runs the code, compares its numbers with the published ones and reports what matched, so the researcher spends time judging the evidence instead of typing commands.
Where you'll meet it
- Long research and engineering tasks run by an AI agent.
- Coding tools such as Codex and Kimi Code, through its API.
- Studying how people and AI agents share research work.
Related
Sources
- Docsinternlm/Atria-Dawn, Shanghai AI Laboratory (Hugging Face) · read 23 Oct 2026
- PaperAtria Dawn Preview: technical report, Shanghai AI Laboratory (arXiv) · read 23 Oct 2026
Checked by WhatIsAI against the sources above · 23 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.