Atria Dawn Preview
#20 in Open-Source LLMsShanghai Ai Laboratory · preview · since 11. September 2026 · 3× · last seen Sep 17, 2026
Atria Dawn Preview is an agentic language model released by Shanghai Artificial Intelligence Laboratory, built on the 744B-parameter MoE GLM-5.2 foundation (753B parameters per the Hugging Face model card, including additional components). It targets research and engineering scenarios requiring continuous environmental understanding, tool use, and multi-step task completion, and was released quietly on September 11, 2026 with MIT-licensed weights on Hugging Face and ModelScope. The model is text-only (not multimodal), supports local deployment (BF16/FP8 via vLLM or SGLang) as well as hosted API access, though no official per-token price has been published.
Features
| Key Benchmark (%) | DeepSearchQA 96.0%; BrowseComp 92.5%; SWE-bench Pro 59.6% (vendor-reported, not independently verified) |
| Context Window (Tokens) | 256,000 tokens (model card); config.json technically lists up to 1,048,576 positions |
| License | MIT License (code and model weights) |
| Multimodality | No – text-only model, does not accept image input |
| Platform | Hugging Face, ModelScope, GitHub; local deployment via vLLM (≥0.23.0) or SGLang (≥0.5.13.post1); hosted API at api.atria-asi.ai |
| Release Date | September 11, 2026 (Hugging Face repo live; FP8 checkpoint Sept 12; technical report Sept 14/15, 2026) |