CUA-Master-27B Hits 80.2 on OSWorld, Beats UI-Mate by 25 Pts

A research collaboration between CUHK MMLab, Tsinghua University, HKUST(GZ) and JD Explore Academy has released CUA-Master-27B, a 27-billion-parameter computer-use agent that scores 80.2 on OSWorld-Verified and trails only GPT-6 Sol and Claude Opus 5 on the hardest cross-platform benchmark. The team posted the model and paper to Hugging Face on November 3, 2026, alongside claims of double-digit gains over the closest same-size competitor across four evaluation tracks.

Benchmark Performance and Competitive Standing

CUA-Master-27B records 80.2 on OSWorld-Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam. On the final benchmark, which the authors describe as the most challenging of the four, the 27B model sits behind only two frontier-scale proprietary systems in the paper's evaluation suite. Against UI-Mate-27B, the nearest open specialist at the same parameter count, the team reports gains of 2.6, 12.9, 25.6 and 11.8 points respectively across the four tracks.

The authors also claim competitive scores against Qwen-CUA-397B-A17B, a model with substantially more parameters, though the paper does not break out per-benchmark comparisons or per-platform results for Windows, macOS and cross-platform task sets.

Nested Double-Loop Training Framework

The model's training pipeline combines two recursive loops. An outer world-data-model flywheel iteratively refines the agent by generating real-world task distributions, collecting interaction trajectories and updating model weights. An inner loop mines seed tasks from YouTube, TikTok and Bilibili video tutorials, then treats those tasks as composable primitives for synthetic task generation. The team ran three supervised fine-tuning iterations stacking 63,000, 38,000 and 14,000 tasks to reach roughly 115,000 total SFT samples. A final reinforcement-learning stage used approximately 12,000 high-quality tasks selected by the inner flywheel.

Training ran on 16 nodes of eight NVIDIA H200-141GB GPUs each. The base model is Qwen-3.5-27B, a Chinese-English bilingual foundation model released earlier in 2026.

Release Date and Availability

CUA-Master-27B was officially released on November 3, 2026. The paper, weights and evaluation code appeared on Hugging Face the same day. The release includes the 27B checkpoint and an ablation suite showing per-iteration SFT gains and RL policy improvements.

Evaluation Protocol and Baseline Comparisons

All baseline scores reported in the paper come from the authors' own re-evaluation runs under a consistent benchmark protocol, not vendor-published numbers. That includes the GPT-6 Sol and Claude Opus 5 comparisons on Agents' Last Exam. The team writes that this ensures fair model-to-model apples-to-apples measurement, but it also means the headline claims rest on internal runs rather than independent third-party validation.

The paper does not specify whether the baselines used official APIs, local inference or a mix. It also does not detail prompt engineering strategies applied to each baseline, a factor that can shift computer-use scores by several points in either direction.

What This Means

CUA-Master-27B demonstrates that a well-structured data generation loop can push a 27B open-weight agent into the top performance tier on multiple computer-use benchmarks. The nested flywheel approach, which mines real-world video tutorials and synthesizes compositional tasks, offers a replicable blueprint for labs without access to massive proprietary task datasets. The model's competitive standing against 397B-scale agents suggests that recursive data curation and multi-stage SFT plus RL can partly substitute for raw parameter scale in agentic workflows, though the lack of independent baseline runs leaves some uncertainty around the exact margin of improvement.