The CUA-Master-27B computer use agent, published on November 2, 2026 in the CUA-Master paper, is the better open pick for mixed GUI and command-line work: it scores 43.9 on WeaveBench and 25.0 on Agents' Last Exam, ahead of Qwen-CUA-397B's 35.1 and 21.1. Qwen-CUA-397B still wins on pure GUI tests, with 84.0 on OSWorld Verified against CUA-Master's 80.2, per the same paper. Pick CUA-Master-27B for agent work that needs a terminal. Pick the 397B Qwen model when the task stays GUI-only and hardware is no constraint.
TL;DR
- An open-weight computer-use model built on Qwen-3.5-27B, trained by iterative SFT and reinforcement learning (paper).
- It scores 80.2 on OSWorld Verified and 36.9 on OSWorld 2.0, both GUI-only (paper).
- Its real edge is hybrid GUI+CLI work: 43.9 on WeaveBench and 25.0 on Agents' Last Exam (paper).
- It beats the closest same-size open specialist, UI-Mate-27B, on all four reported benchmarks.
- No licence is declared on the weights or the code, and no quantised build exists, so treat it as research-grade for now.
- Last verified: November 4, 2026.
Which is better, CUA-Master-27B or Qwen-CUA-397B?
It depends on the task type. The paper's comparison table gives these results (OSWorld Verified, OSWorld 2.0, WeaveBench, Agents' Last Exam), all from the CUA-Master paper:
| Model | OSWorld Verified | OSWorld 2.0 | WeaveBench | Agents' Last Exam |
|---|---|---|---|---|
| CUA-Master-27B | 80.2 | 36.9 | 43.9 | 25.0 |
| Qwen-CUA-397B-A17B | 84.0 | 43.6 | 35.1 | 21.1 |
| UI-Mate-27B | 77.6 | 24.0 | 18.3 | 13.2 |
| Qwen-3.5-27B (base) | 54.6 | 0.8 | 13.5 | 1.3 |
The 397B Qwen model leads both GUI-only benchmarks; CUA-Master leads both hybrid benchmarks, which mix mouse-and-screen actions with shell commands. Against UI-Mate-27B, the closest same-size model, CUA-Master gains 2.6, 12.9, 25.6 and 11.8 points across the four tests (paper).
One more nuance: the generalist Qwen-3.8-27B reportedly scores 81.2 on OSWorld Verified and 36.5 on OSWorld 2.0 (paper), so on GUI-only tests CUA-Master's lead over a strong general model is thin. The benefit shows up when the CLI is involved. That positioning is our own read: We captured live top-10 organic results for 7 target queries to see who actually occupies the page rather than who claims to (n=7, captured 2026-11-04).
What makes the CUA-Master-27B computer use agent different?
The training pipeline matters more than the model. The authors describe a nested double loop (paper): an outer data-and-model flywheel where current weaknesses steer what training data gets built next, and an inner loop with two stages:
- Open-world seeding. Workflows are mined from public instructional videos and rebuilt into trajectories from spoken-step clauses plus on-screen visual changes.
- Autocatalytic propagation. Seed tasks break into reusable ability primitives, then recombine into new tasks with controlled complexity and composable verifiers.
Trajectories are collected in sandboxed environments with snapshot-and-rollback retries; RL then trains on outcome rewards from those verifiers.
The flywheel ablation shows the effect: 40 primitives across 3 apps reached about 50% success against a 14.6% baseline; 50 primitives across 8 apps reached 18.7% against 0.0% (paper). Our reading: data construction, not parameter count, moved a 27B model into the top tier — that is our interpretation, not a stated author claim.
How fast is it compared with similar models?
The authors report fewer steps per task and 24 to 36% less runtime than comparable models (paper). On OSWorld 2.0 it scores 36.9 against Qwen-3.8-27B's 36.5 using 24% less runtime; on WeaveBench it matches GPT-5.6-Luna with 28% less runtime.
On Agents' Last Exam the authors say it lags only GPT-5.6-Sol and Claude-Opus-5 and surpasses all other counterparts (paper). If you are weighing open models against closed ones on speed and capability, our GPT-5.6 Sol speed comparison covers the closed side of that trade.
Can you run CUA-Master-27B on a 12 GB GPU?
No, not today. The Hugging Face repository holds BF16 safetensors for 27B parameters, totalling 54.7 GB across 12 shards, and no GGUF or other quantised build was listed when we checked on November 4, 2026.
This is our arithmetic, not a published figure. A 4-bit build of a 27B model needs roughly 14 to 16 GB for weights alone, before KV cache and vision components — already over a 12 GB card. A 24 GB RTX 4090 could manage 4-bit with some CPU offload, tightly. Published BF16 means a datacentre GPU or several smaller ones. Until a community quantisation appears, expect a rented high-memory card or patience.
Is CUA-Master-27B safe and licensed for production use?
Not yet. The weights repository lists no licence tag, and the GitHub repository has no LICENSE file as of November 4, 2026. The base model, Qwen3.5-27B, is tagged Apache-2.0 on Hugging Face, but that does not automatically extend to a derivative. As published, the weights are available, not licence-clean.
The model card's own safety note is worth following: because actions may modify files or execute commands, use a sandboxed environment and verify outcomes before granting broader permissions (model card). Third-party summaries of the release, including AI Weekly, report the same training scale, about 115K SFT and 12K RL samples. The authors did the headline evaluations themselves. We found no independent third-party replication of the scores, and the mix of API and local inference used for the baselines is not specified; the independent outlet emergent.sh flagged the same limitation on November 3, 2026.
How do builders get started with the CUA-Master repository?
The GitHub repository exposes the pipeline as CLIs, in practice this order:
- Install the extras:
pip install -e ".[seeding,annotator,collection,dev]". - Copy
.env.exampleto.env; add annotator API keys and sandbox endpoints. - Implement the
SandboxProviderinterface for your VM orchestration. - Run harvesting, transcription, scoring and reconstruction for seed tasks.
- Run
cuamaster-synthesize,cuamaster-collect,cuamaster-dataset.
Inference-only users can skip the pipeline: serve the weights behind an OpenAI-compatible endpoint and use the repo's validate_action action space, which expects click coordinates in a 0 to 1000 range. Scripts exporting SFT data in ms-swift format are included for fine-tuning.
For the day-to-day picture, see our ChatGPT computer-use agents guide and background computer use with Hermes Agent. For code-focused agents, start with the best AI agent for coding in 2026. For the conceptual split between acting agents and text generators, see agentic AI vs generative AI.
Verdict: who should use CUA-Master-27B?
Choose CUA-Master-27B if you are researching or prototyping agents that mix GUI and terminal work, you can host 27B hardware, and an undeclared licence is acceptable. It holds the strongest hybrid scores among the open models in the paper's table, with runtime numbers suggesting lower per-task cost than same-size peers. Choose Qwen-CUA-397B-A17B for GUI-only work if you have the hardware, since it wins both GUI benchmarks. Avoid either for production until licensing is clear and independent evaluations exist.
FAQ
Q: What is CUA-Master-27B?
A: An open-weight computer-use agent built on Qwen-3.5-27B, trained by iterative SFT and RL, described in the CUA-Master paper released November 2, 2026.
Q: What does CUA-Master-27B score on OSWorld?
A: It scores 80.2 on OSWorld Verified and 36.9 on OSWorld 2.0, both GUI-only, according to the paper. Qwen-CUA-397B-A17B scores higher on both, at 84.0 and 43.6.
Q: Is CUA-Master-27B better than GPT-5.6-Sol and Claude-Opus-5?
A: No. On Agents' Last Exam the authors say it lags only GPT-5.6-Sol and Claude-Opus-5, scoring 25.0, as reported in the paper. It is the strongest open model in their table on that test. AI Weekly independently reported the same four scores on November 3, 2026, and CCTest walked through the method the same day.
Q: Is CUA-Master-27B free for commercial use?
A: Unclear. Neither the Hugging Face weights nor the GitHub repository declares a licence as of November 4, 2026; ask the authors before commercial use.
Q: Can I run CUA-Master-27B locally?
A: Only on large hardware for now. The published weights total 54.7 GB in BF16 on Hugging Face, and no quantised build existed when we checked. In our estimate, a 12 GB GPU is not enough.
Corrections log
4 November 2026: first published after verifying all scores and hardware figures against the paper, the model card and the GitHub repository on 2026-11-04.
This article is AI-assisted: researched, drafted and fact-checked by our AI pipeline with every figure linked to a primary source, and reviewed by our editorial team. Learn more in How We Work.
Top comments (0)