Shanghai AI Lab Released a 744B Agentic Model Built on GLM 5.2 Under an MIT License

Shanghai Artificial Intelligence Laboratory published the weights of Atria Dawn Preview to Hugging Face under an MIT license on September 11, 2026, and released a technical report with 143 authors as arXiv 2609.15818 on September 14. The model is an agentic model built on GLM 5.2, a 744-billion-parameter MoE foundation model, it carries a 256K token context window, and it accepts text input only. In the 16-benchmark comparison table published in the model card, Atria Dawn Preview holds the top score on five rows with DeepSearchQA 96.0, BrowseComp 92.5, BFCL v4 77.0, AutomationBench 53.8 and CyberGym 86.5, while on SWE-bench Pro in the same table it scores 59.6 against Claude Opus 5's 74.7, a gap of 15.1 points. ASAP separates what is a new result in this release from what is inherited.

The base model field says GLM 5.2

The opening paragraph of the model card states that Atria Dawn Preview is built on GLM 5.2, a 744-billion-parameter MoE foundation model. The architecture tag on the Hugging Face repository reads glm_moe_dsa, and the local deployment section links to the GLM 5.2 page of the SGLang cookbook and the zai-org/GLM-5.2 page of the vLLM recipes. What Shanghai Artificial Intelligence Laboratory published here is therefore not a newly pretrained foundation model but an agentic post-training layer placed on a base another organization pretrained.

That distinction changes how the numbers should be read. The 744B figure says nothing about this lab's pretraining investment. What the release argues instead is that teaching verifiable experience to an open base model can reach frontier agent territory without pretraining capital. The method the technical report puts forward is not a new architecture either but a Verifiable Experience Pipeline, a structure that connects tool-mediated interactions to executable environments and externally verified outcomes.

It is also evidence that the division of labor inside China's open model ecosystem has moved one step further. Pretraining sits with Zhipu's GLM line, agentic post-training sits with a national lab, and the result goes back out under MIT. The fact that such a handoff works across organizational boundaries at all is a path that has no equivalent in the closed-model camp.

The five first places cluster in search and tool calling

The comparison table places DeepSeek V4 Pro 0813, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 sol and Claude Opus 5 as the reference set and groups 16 benchmarks into five categories. The five rows where Atria Dawn Preview leads share a clear character. Two sit in Discovery, DeepSearchQA at 96.0 (Kimi K3 95.9, GLM 5.3 94.7, GPT 5.6 93.2) and BrowseComp at 92.5 (GPT 5.6 92.2, Kimi K3 91.2, Opus 5 90.8). Two sit in Tool Use, BFCL v4 at 77.0 (GLM 5.3 74.1, DeepSeek V4 Pro 71.4) and AutomationBench at 53.8 (Qwen 3.8 Max 49.7, Opus 5 49.4). The last is CyberGym at 86.5 in Cybersecurity (GLM 5.3 84.5, GPT 5.6 83.6).

The size of each lead differs sharply. The gap between first and second is 0.1 points on DeepSearchQA and 0.3 points on BrowseComp, so reading those as ties is the accurate move. AutomationBench, by contrast, leads by 4.1 points, BFCL v4 by 2.9 and CyberGym by 2.0. The firm advantage in this model therefore lies less in raw search accuracy than in the stretch where it calls tools to automate a procedure, plus vulnerability analysis.

Rows split even inside a single category. On WideSearch in Discovery the model posts 81.9 behind GPT 5.6's 83.3 and GLM 5.3's 82.7, and on DeepResearch Bench II it posts 51.1 behind Opus 5's 54.1 and GLM 5.3's 52.7. In Tool Use, SkillsBench at 66.4 trails Qwen 3.8 Max's 66.7 by 0.3 points and τ³-Bench Banking at 41.2 trails Qwen 3.8 Max's 55.2 by 14.0 points. There is effectively no category this model wins as a whole.

On code and delivered work the distance to Opus 5 is intact

Across the seven rows of the Creation and Delivery categories, Atria Dawn Preview never takes first place, and Claude Opus 5 holds the best score on six of those seven. SWE-bench Pro is 59.6 against 74.7, a 15.1-point gap. Terminal-Bench 2.1 is 78.3 against 90.2, 11.9 points. JobBench is 50.3 against 68.0, 17.7 points. GDPval is 1583 against 1768, and on that row Qwen 3.8 Max at 1722, GPT 5.6 at 1682 and GLM 5.3 at 1667 also sit ahead of Atria.

That contrast fixes the real coordinates of this release. The front half, gathering evidence and calling tools, has caught up to frontier level, while the back half, editing a codebase until it passes and carrying documents and reports to completion, remains a real distance away. The model card names four axes of its own, Discovery, Creation, Delivery and Cybersecurity, and the strong ones are Discovery and Cybersecurity while the weak ones are Creation and Delivery.

The Workspace rows are the near-miss exception. Workspace-Bench is 65.0 against 65.8, a 0.8-point gap, and Workspace-Bench-Lite is 68.2 against 70.1, 1.9 points. The pattern is that office-document tasks are close while JobBench and GDPval, which span whole job functions, are not, and evaluating Delivery as a single block hides exactly that difference.

The table also carries a fair number of blanks. DeepSeek V4 Pro and Qwen 3.8 Max are marked unavailable on DeepSearchQA, as are Qwen 3.8 Max, GPT 5.6 and Opus 5 on BFCL v4 and Opus 5 on CyberGym. Two of the five first places, BFCL v4 and CyberGym, were therefore won with strong competitors' scores missing, and the table's own authors marked those gaps.

The technical report spends as much space on 56 people as on the model

Close to half of the abstract of arXiv 2609.15818 is a record of how the model was built rather than of what it scores. The authors analyzed 769 task records from 56 participants together with agent logs, and they write that when participants were asked to evaluate completed tasks under comparable conditions, they rated about one third of the completed AI-assisted tasks as infeasible without AI. The abstract states that agents frequently proposed methods and implemented revisions, while humans retained most final decisions and guided exploration through judgment and feedback.

Attaching a team's own work logs to a model release report is unusual. Benchmark scores speak to a model's capability, while this record speaks to how that capability actually entered a research process. The phrase the abstract uses is a shift from task-level execution to project-level partnership, with human effort concentrating on what is worth pursuing and how evidence should guide research.

The figure still needs careful handling. That one-third value is self-reported by the team that built the model, not the output of a controlled experiment. Participants were rating tasks they had just finished with AI, which leaves room for recency to enter the judgment, and the abstract itself tempers the claim with the word observations. What matters more than the size of the number is that the item was put in the report at all.

The paper is titled Atria Dawn: The Dawn of Agentic Superintelligence, while the closing sentence of its abstract calls for preserving accountable human authority over the risks and direction of continued development. The temperature difference between the title and the ending summarizes this report's character.

Open weights and runnable weights are not the same statement

The distribution terms are unambiguous. Code and weights are both under MIT, an Instruct checkpoint and an FP8-quantized Instruct checkpoint sit on both Hugging Face and ModelScope, and both carry the 256K token context. Local inference requires SGLang v0.5.13.post1 or later, or vLLM v0.23.0 or later, and hosted access is offered through a separate international endpoint at api.atria-asi.ai and a China endpoint.

Yet the Hugging Face API reports cumulative downloads of 711 and 180 likes for this repository as of September 19, 2026. Those numbers show the distance between weights being open and weights being runnable. A 744-billion-parameter MoE stays outside the reach of personal hardware or a small server even in FP8, so for most teams the practical point of contact with this model is the API rather than local deployment.

Teams evaluating it are better off inverting the usual order. Rather than starting with whether self-hosting is feasible, run your own work tasks through the API and measure whether the search and tool-calling stretch actually produces a gain. If the model card's table holds, the likeliest wins are in web evidence gathering and automation pipelines, and keeping an incumbent frontier model for code edits and report production is the reasonable call.

One constraint deserves checking up front. The model card states that Atria-Dawn-Preview accepts text input only, and that attaching an image makes the endpoint return a 400 error saying it is not a multimodal model. Agent workflows that read screenshots to operate a screen do not drop into this model as is.

What the word preview still has to answer for

The first gap is who did the measuring, because the 16-benchmark comparison table was produced by the team that built Atria Dawn Preview and no third-party reproduction is attached to it yet. Whether measurement conditions and scaffolding were identical for the competing models is not stated in the model card, and agent benchmarks are an area where harness configuration moves results substantially.

The second gap is separation from the base model. Because this is post-training on top of GLM 5.2, the table alone cannot say which portion of the five first places comes from the post-training and which from the base. GLM 5.3 appears among the comparisons, but this model's base is 5.2, so it is not a direct control either. Answering that question requires a matched comparison starting from the same base.

The third gap is the word preview itself. The model card describes this as a preview of a new-generation agentic model, and the repository still holds a commented-out citation block and commented-out local-run instructions. The evaluation figure carries a to-do note asking for chart annotations. Reading this as a release where the weights shipped ahead of the documentation is the accurate framing.

In short, the value of this release sits in the path rather than in five top scores. Agentic post-training alone, laid on an open base model, brought the search and tool-calling stretch up to frontier level, and the result went back out under MIT. The next checkpoints are third-party reproduction and a matched-base comparison, and whether the double-digit gaps on code edits and delivered work narrow in the next version is what decides the limits of this path.

Sources: Atria-Dawn-Preview model card (Shanghai Artificial Intelligence Laboratory, Hugging Face, published September 11, 2026) · Atria Dawn: The Dawn of Agentic Superintelligence (arXiv:2609.15818, September 14, 2026)