Shanghai Artificial Intelligence Laboratory released the full version of Atria Dawn on October 23, 2026. It is an agentic model aimed at research and engineering work: deep research, software building, document delivery and, in authorised environments, security analysis and re-validation. The weights are downloadable from Hugging Face and ModelScope, in full precision and in a separate FP8 repository, and the model card declares them MIT-licensed. A hosted API is live alongside local deployment through SGLang and vLLM, with OpenAI- and Anthropic-compatible surfaces.

The headline number is 744 billion parameters. Treat it carefully. According to Shanghai AI Lab's model card, Atria Dawn is built on a 744-billion-parameter mixture-of-experts GLM-5.2 foundation. The card does not explain the technical or organisational relationship between this release and the GLM-5.2 foundation, and nothing in this article assumes one. Arithmetic on the published configuration — 78 layers, 256 routed experts per MoE layer — comes out at roughly the same total for the released weights, so the 744B describes the download, not just the architecture it derives from.

What the card does not give is the active-parameter count: how many parameters fire per token. That count is an important input to serving-cost estimates, alongside total weight memory, quantisation, batching, expert routing, context length and hardware. The configuration does show the routing — 8 of 256 routed experts, plus one shared expert, per MoE layer. ALFW.AI's calculation from those figures puts the active count at roughly 40 billion parameters. That is our arithmetic, with the usual estimation caveats, not a number the lab publishes. Anyone comparing serving costs should treat the lab's silence, and our estimate, accordingly.

The table that argues against itself

The evaluation table is the strongest and weakest part of the release at once. On the lab's own figures, Atria Dawn tops the lab's selected comparison set on AutomationBench (54.7), CyberGym (86.8), SkillsBench (67.8) and WideSearch (84.6), and posts 96.5 on DeepSearchQA and 91.3 on BrowseComp, the last of those lab-reported like everything else in the table. Those are search, tool-use and automation results, and they fit the model's stated purpose.

In Shanghai AI Lab's published evaluation table, Atria Dawn scores 61.7 on SWE-bench Pro against 76.5 for Claude Opus 5; 50.0 against 79.9 on NL2Repo-Bench; and 49.3 against 72.6 on DeepSWE. The lab also reports 81.0 versus 90.2 on Terminal-Bench 2.1, 17.3 against 41.0 on Harbor-Index, and GDPval at 1594.2 against 1714.0 for Opus 5 and 1710.9 for Qwen 3.8 Max. These are the lab's reported scores for both its model and its competitors, not independently reproduced results.

The lab's table also places Atria Dawn first among the models it compares on Harvey LAB, at 27.5; that ranking does not establish independent performance or practical legal-task reliability.

The table is useful because it includes weaker results as well as stronger ones. That makes the lab's stated performance profile easier to scrutinise, but it does not establish that the evaluations were conducted on comparable terms or that the results will reproduce independently. The card adds to the caution: it says the table matches the lab's official website and states no benchmark versions, configurations or evaluation dates, so even the comparisons it publishes cannot be re-run from what the release discloses.

The licence, checked

The card's licence section states it plainly: “The code and model weights in this repository are released under the MIT License.” The release badge says MIT too, and this article found no additional terms or restrictions attached to the weights anywhere in the repositories. The paperwork, though, is untidy. The LICENSE file that the licence statement links to is not in the main Hugging Face repository — the link returns a 404 — nor in the FP8 repository, and neither full-release repository carries licence metadata. The September Preview repository's metadata does declare MIT, and the project's GitHub repository contains the full MIT licence text. On the evidence available, MIT is the lab's declared licence for code and weights; a careful adopter will still want the licence file itself in the download before building on it. Separately, MIT-licensed weights are not a fully open training pipeline: the repositories publish weights, configuration and tokenizer files, not training data or a training recipe.

Two limits stated plainly

The card is direct about input. Atria Dawn accepts text only; it is not a multimodal model, and the card documents the hosted endpoint rejecting image input with a 400 error — “Atria-Dawn is not a multimodal model”. The lengthy integration sections for Codex, Claude Code and Kimi Code exist largely to stop those tools from sending it pictures and PDFs. For a model pitched at office and research delivery, that is a real constraint: screenshots, scanned documents and charts have to be turned into text somewhere else first.

Scale is the other constraint. A 744B MoE, even in FP8, is not a laptop model and not a single consumer GPU model. The practical audience is labs and companies with serious inference infrastructure, or users of the hosted API — who should verify pricing, rate limits, data retention and processing location with the provider, because the model card prices none of it out.

For readers tracking open weights, the release still matters, and the correct test is narrower than the launch language about agentic superintelligence. Start independent evaluations with the search, tool-use and automation workloads on which the lab reports strong results. Use representative tasks, compare against alternatives under matched conditions, and measure latency, cost and reliability before making deployment decisions — and wait for the lab to state an active-parameter count and file its licence paperwork before treating either the cost or the openness as fully settled.

Sources

  1. [1] InternLM / Shanghai Artificial Intelligence Laboratory, “Atria Dawn” model card and repository files, Hugging Face (introduction and GLM-5.2 foundation statement; model downloads table; full evaluation table; deployment via SGLang and vLLM; Codex, Claude Code and Kimi Code integration sections, including the documented 400 rejection of image input; licence section; config.json; file list and commit history — all read in full for this article, 23 October 2026)Read source
  2. [2] InternLM, “Atria-Dawn-FP8” repository, Hugging Face (separate FP8 download; file list checked — no LICENSE file present)Read source
  3. [3] InternLM, “Atria-Dawn-Preview” repository, Hugging Face (September 2026 Preview at 256K context; its card metadata declares the MIT licence)Read source
  4. [4] Atria ASI project repository on GitHub (contains the MIT LICENSE file for the project)Read source
  5. [5] The Deploy Log, “What shipped in AI, week of October 19, 2026” (independent shipping log entry recording the 23 October release on Hugging Face and ModelScope; release-date corroboration)Read source