Shanghai AI Lab ships Atria Dawn full release: 744B MoE GLM-5.2 agentic model, context doubled to 512K, MIT weights

Background and the problem

Shanghai AI Laboratory open-sourced Atria Dawn Preview in mid-September as a long-horizon agent model meant to turn research questions into verifiable results. Open-weight agent models have tended to struggle with context length and with keeping multi-step tool use stable over long experiments, large reading loads or full engineering runs. The full release is now on Hugging Face (internlm/Atria-Dawn, repo created October 23) and ModelScope, which announced it on X on October 27. The license stays MIT.

Architecture and how it works

Per the official model card, Atria Dawn sits on the 744B-parameter MoE GLM-5.2 and covers four areas: retrieval and deep research, building software and apps, turning documents and data into reports, and security analysis in authorized environments. It is built around a full loop of analysis, planning, tool use, coding, running experiments, reading results and recovering from failures. Context rises from 256K to 512K, input is text only, and an FP8 build ships alongside. The technical report is arXiv 2609.15818.

Benchmarks

All figures are vendor-reported. The full release posts the top score in its table on AutomationBench (54.7), CyberGym (86.8), SkillsBench (67.8), WideSearch (84.6) and BFCL V4 (78.4), against DeepSeek V4 Pro, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 sol and Claude Opus 5. Coding is clearly weaker: SWE-bench Pro 61.7 versus 76.5 for Claude Opus 5, Terminal-Bench 2.1 81.0 versus 90.2, and NL2Repo-Bench 50.0, slightly below the Preview's 51.4. No independent third-party results have been published yet.

Getting started

For local serving the card recommends SGLang v0.5.13.post1+ or vLLM v0.23.0+ using the GLM-5.2 recipes; at over 700B parameters this needs a multi-GPU server, and the FP8 build eases memory. The hosted API is OpenAI-compatible at discovery-api.intern-ai.org.cn/v1 with model Atria-Dawn, and the card includes a Codex config.toml custom-provider snippet. It looks best suited to research automation, office deliverables and security validation agents; for pure coding, closed frontier models still lead.