CUA-Master-27B Scores 80.2 on OSWorld Verified from CUHK, THU, JD

A 27B-parameter open-weight agent now sits inside the top tier on four computer-use benchmarks. The paper, posted to Hugging Face on 3 November by a team from MMLab CUHK, Tsinghua, HKUST(GZ) and JD Explore Academy, reports that CUA-Master-27B hits 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam. On the hardest of those, the authors write that "it only lags behind GPT-5.6-Sol and Claude-Opus-5, and surpasses all other counterparts."

The model is initialized from Qwen-3.5-27B and run through a "nested double-loop recursive improvement framework": an outer world-data-model flywheel plus an inner data flywheel that mines seed tasks from YouTube, TikTok and Bilibili tutorials, then uses them as composable primitives for synthetic task generation. Three SFT iterations stack 63K, 38K and 14K tasks respectively, for about 115K total; a final RL stage uses roughly 12K high-quality tasks. Training ran on 16 nodes of 8 NVIDIA H200-141GB GPUs.

Against UI-Mate-27B, the closest same-size specialist, the authors claim gains of 2.6, 12.9, 25.6 and 11.8 points across the four benchmarks. They also report competitive scores against Qwen-CUA-397B-A17B despite "substantially smaller parameters."

The caveats sit in the method section. All baselines were re-evaluated by the authors "under a consistent benchmark protocol" — not vendor-run — so the GPT-5.6 and Claude-Opus-5 comparisons rest on the paper's own runs. The passage does not break down per-benchmark numbers for Windows, macOS or the cross-platform suite it flags.

The release lands in a dense week of same-size CUA work on our radar, including NorthCUA Pro earlier on 3 November, part of 526 agents stories we have tracked in the last 90 days.