Shanghai AI Lab open-weights Atria Dawn Preview, a 744B MoE agent built on GLM-5.2
MIT weights on Hugging Face and ModelScope, 256K context, text-only. Vendor card: DeepSearchQA 96.0 and CyberGym 86.5; SWE-bench Pro 59.6 against Claude Opus 5’s 74.7. Recipes for SGLang, vLLM, Codex, Claude Code, Kimi.
Deep dive2 minIn-depth
Maya Chen2 min read
- Shanghai AI Lab
- Atria
- Open weights
- GLM-5.2
- Agents

Photo: Hugging Face / Shanghai AI LaboratorySource page
The Shanghai Artificial Intelligence Laboratory has posted Atria Dawn Preview, an agentic instruct model built on Z.ai’s 744-billion-parameter mixture-of-experts GLM-5.2, on Hugging Face and ModelScope. The card lists a 256,000-token context, text-only input — it is not multimodal — and MIT-licensed weights in BF16 and FP8. A hosted endpoint sits at api.atria-asi.ai. An accompanying paper, “Atria Dawn: The Dawn of Agentic Superintelligence,” is on arXiv as 2609.15818. The lab says the model is aimed at research and engineering work that needs tools, a long environment loop, and results you can re-run: discovery, software creation, office-style delivery, and authorised cybersecurity.
What the lab calls a Verifiable Experience Pipeline is the training claim: agent skills are pushed into the weights so the model can take a question through analysis, tools, code, experiment, and recovery without a separate skill pack at inference. That is a vendor description. The useful part of the card is the score table, which the lab itself ran, and the places it does not win.
Strong on search and cyber; behind Opus on SWE-bench Pro
On the published board, Atria Dawn Preview leads several discovery and security rows the lab chose. DeepSearchQA is 96.0, ahead of Kimi K3 at 95.9 and GPT-5.6 Sol at 93.2. BrowseComp is 92.5, a hair over Sol’s 92.2 and Claude Opus 5’s 90.8. CyberGym is 86.5, above DeepSeek V4 Pro 0813, GLM 5.3 and Sol. BFCL v4, a tool-use set, is 77.0. Those are the numbers the card bolds. They are also the numbers a lab gets to pick the columns for.
The creation row is less kind. SWE-bench Pro is 59.6 against Claude Opus 5’s 74.7 — a fifteen-point gap on the software-engineering set most buyers actually care about. Terminal-Bench 2.1 is 78.3 against Opus at 90.2 and Qwen 3.8 Max at 89.3. MLE-bench Lite is competitive at 86.2, still behind Sol and Opus. Delivery scores sit in the pack; JobBench is 50.3 against Opus at 68.0. Read that as an open-weight agent that the lab has tuned for search, browsing and authorised cyber work, and that does not yet match the closed coding frontier on the bench it published. Independent reproductions are not on the card.
Text-only, and the harnesses that assume otherwise
Local serving is documented for SGLang 0.5.13.post1 or later and vLLM 0.23.0 or later. The longer recipes are for coding harnesses that assume every model can see images. Codex, the card says, will attach a screenshot from a paste or `-i` and the Atria endpoint will reject it with a 400: not a multimodal model. The fix is a model-catalog file that sets `input_modalities` to text only, plus Codex CLI 0.154.0 or later. Claude Code gets a PreToolUse hook that blocks Read on PDFs and images. Kimi Code 0.43.1 and later talks to the same host over Chat Completions; the sample config has to declare tool use and thinking, and cap `max_output_size` at 65,536 or the API rejects the call. Those are integration notes, not benchmarks. They are also why a 744B open-weight agent is not a drop-in for a Claude or GPT slot this week. The weights are MIT. The harness still has to be told the model is blind.


