Two researchers on opposite sides of the Pacific run the same experiment overnight: hand a model a messy research question and see whether it can plan, search code, run the analysis and write up what went wrong. One of them sends every intermediate step to a hosted service and hopes the terms do not change. The other downloads weights once and answers to nobody. Until recently, the second researcher had to accept a much weaker model.
Shanghai AI Laboratory (上海人工智能实验室) has been working to erase that trade-off. After a preview release in September, it published the complete weights of Atria Dawn, an agentic large model (大模型) designed to turn open-ended problems into executable, verifiable results — under an MIT licence, with no usage restrictions beyond attribution.
Key takeaways
- Atria Dawn is built on a 744B-parameter mixture-of-experts (混合专家) GLM-5.2 foundation and supports a 512K-token context, double the preview's 256K.
- The instruct and FP8 checkpoints are on Hugging Face and ModelScope under MIT, one of the most permissive licences available for a model of this size.
- Vendor-reported benchmarks put it at the top of its own comparison table for AutomationBench, CyberGym, SkillsBench, WideSearch and BFCL V4, while trailing Claude Opus 5 on pure coding benchmarks.
- Serving requires a multi-GPU server; the weights alone weigh roughly 753 GB.
What an "agentic" model actually promises
The word agent gets attached to almost everything now, so it helps to be specific about what this model is trained to close. The model card describes a full loop: analyse a problem, design an approach, call tools, write and run code, execute experiments, read the results and recover from failures. Four capability dimensions organise the work — discovery (evidence retrieval and deep research), creation (software, applications, games, data visualisations, machine-learning systems), delivery (turning documents and data into reports and presentations) and cybersecurity (validating vulnerabilities, applying fixes and re-verifying them in authorised environments).
That framing matters because long-horizon agent workloads stress models in ways chat benchmarks never measure. Tool-heavy trajectories stretch the context window with prefill-heavy inputs; multi-step tasks punish models that lose the thread after a few dozen turns. Atria Dawn's headline architecture choice — a very large mixture-of-experts base paired with a doubled 512K context — is aimed squarely at that workload shape. Input is text-only, which rules out screenshot-driven workflows but simplifies deployment.
The numbers, and who produced them
Every benchmark figure below is vendor-reported. In the laboratory's own table, Atria Dawn scores 54.7 on AutomationBench, 86.8 on CyberGym, 67.8 on SkillsBench, 84.6 on WideSearch and 78.4 on BFCL V4 — the best row scores in its comparison, which includes DeepSeek V4 Pro, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 sol and Claude Opus 5. On MLE-bench Lite it reports 88.2.
Coding is the honest weak spot. The model reports 61.7 on SWE-bench Pro against 76.5 for Claude Opus 5, with Qwen 3.8 Max at 65.3 and Kimi K3 at 62.0; Terminal-Bench 2.1 stands at 81.0 versus 90.2 for Opus 5. Repository-generation work (NL2Repo-Bench) trails furthest at 50.0. The laboratory did not sweep these under the rug — the same card that claims the automation wins also publishes the coding losses, which is more transparency than most release notes manage.
Because the weights are public, the vendor-reported caveat has an expiry date. Independent evaluation groups can now run SWE-bench Pro and CyberGym themselves on the full 512K checkpoints rather than trusting the laboratory's table. No independent results have been published yet; that replication step is the real test of the release.
Why the licence line matters more than the leaderboard
The MIT licence is arguably the most consequential line in the entire model card. It permits anyone to use, modify, redistribute and commercialise the weights with little more than an attribution notice. Contrast that with recent open-weight releases that ship under non-commercial community licences, which let researchers tinker but push companies back toward hosted APIs. A 744B-parameter model in the same permissive bracket as a small fine-tune removes an entire category of legal negotiation.
The practical trade-off is hardware. At roughly 753 GB of BF16 weights across 353 files, serving the model locally needs a multi-GPU node; the FP8 build eases memory pressure, and the model card names SGLang v0.5.13.post1 or newer and vLLM v0.23.0 or newer as supported runtimes, pointing to the GLM-5.2 deployment guides. Teams that cannot run it locally can use the hosted OpenAI-compatible endpoint with the model id Atria-Dawn, and the card includes custom-provider configurations for coding agents such as Codex and Claude Code.
For readers outside China, this release is part of a pattern worth tracking: frontier-scale agentic models moving into fully open licences, and the competitive boundary shifting from raw capability toward trust and deployment terms. According to Li An, chief scientist at BrainNet (脑机网), the question the next year will settle is not who publishes the biggest parameter count but who can be audited — weights that organisations can inspect, run inside their own walls and hold accountable when they fail.
Honest limitations
- All benchmark results are vendor-reported, and no independent third-party evaluation of the full release has been published yet.
- The model trails frontier closed systems on coding benchmarks — SWE-bench Pro 61.7 versus 76.5 — and repository-scale code generation is its weakest measured area.
- It is text-only, so workflows that depend on reading screenshots or interfaces are out of scope.
- Local deployment requires serious hardware; the 753 GB weight footprint is far beyond a workstation, and the FP8 variant helps but does not make it consumer-accessible.
- CyberGym's 86.8 measures vulnerability work in authorised test environments; real-world security performance is unproven, and downloadable weights cannot be patched or withdrawn once copied.
What readers can do now
- If you operate evaluation infrastructure, run SWE-bench Pro and AutomationBench on the public checkpoints and publish your setup — independent numbers are the release's missing piece.
- If your organisation needs an auditable agent for research automation or document delivery, test the hosted endpoint first, then price the multi-GPU node you would need for self-hosting under MIT terms.
- If you maintain a coding-agent stack, note that the model card documents custom-provider configuration for Codex and Claude Code; text-only input means you can skip vision plumbing entirely.