open-source

Shanghai AI Lab posts full Atria Dawn weights under MIT, with a 512K context

Atria Dawn is an agent model built on the 744B-parameter GLM-5.2 mixture-of-experts base. The full release and an FP8 version went up on Hugging Face on October 23, doubling the preview's 256K context. Code and weights are MIT licensed.

Written by Mara at Sig · Oct 25, 2026

The model is text only. The lab's own table (self-reported) has it ahead on agent and tool benchmarks like AutomationBench (54.7) and BFCL V4 (78.4), but well behind Claude Opus 5 on coding: 61.7 vs 76.5 on SWE-bench Pro and 50.0 vs 79.9 on NL2Repo.

You can run it locally with SGLang v0.5.13.post1 or later, or vLLM v0.23.0 or later. There's also a hosted API with OpenAI- and Anthropic-compatible endpoints. The model card includes setup for Codex CLI, Claude Code and Kimi Code. Hardware requirements aren't published, and at 744B parameters this is a multi-GPU server job, not a laptop one.

One gotcha from the card: Codex assumes every model takes images, so pasting a screenshot gets a 400 error. The card's fix is a model catalog file with input_modalities set to text.

Why it matters

MIT is about as permissive as licenses get, so you can fine-tune and ship this commercially with no revenue cap or credit line. That's a different deal from MiniMax's M3.1 Flash community license this week.

Try this

To try it in Codex CLI without hosting it, add the custom provider from the model card to ~/.codex/config.toml (base_url https://discovery-api.intern-ai.org.cn/v1, wire_api responses, model Atria-Dawn) and create the catalog file with "input_modalities": ["text"], so pasted images don't fail with a 400. The card says this needs Codex CLI 0.154.0 or later.

Sources

Back to all news Get Sig Weekly every Monday