Is CUA-Master-27B state of the art?Pending benchmarks.Announced Nov 2, 2026; no independent results yet.
A 27B computer-use agent model from CUHK MMLab and collaborators, built on Qwen3.5-27B for GUI-only and hybrid GUI + CLI tasks and trained with iterative SFT and RL on verified trajectories mined from instructional videos and synthesized tasks. Weights on Hugging Face; no license stated.
The details so far
- Lab
- CUHK MMLab
- Announced
- Nov 2, 2026
- API
- Not yet · None announced
- Open weights
- Published · Hugging Face yxl66666/CUAMaster-27B (an author's account, linked from the paper), BF16 safetensors, not gated
- Parameters
- 27B (27,356,728,560 stored BF16 parameters)
- Type
- Computer-use agent model for GUI-only and hybrid GUI + CLI workflows
- Base model
- Qwen3.5-27B
- Training
- Iterative SFT and RL (a GRPO-style online RL setup following ScaleCUA) on verified trajectories from tasks mined from instructional videos and from compositional task synthesis
- License
- Not stated on the model card or the code repository
- Code
- github.com/xlyu0106/CUA-Master: data pipeline (video harvesting, trajectory reconstruction, task synthesis, sandboxed trajectory collection) and SFT data preparation
- Paper
- CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements (arXiv 2611.03126)
- Authors
- Xinlei Yu and 17 co-authors from CUHK MMLab, Tsinghua University, HKUST (Guangzhou) and JD Explore Academy
- Requirements
- A computer-use execution environment with GUI and/or CLI actions; the card advises a sandbox
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: OSWorld Verified 80.2 and OSWorld 2.0 36.9 (GUI), WeaveBench 43.9 and Agents' Last Exam 25.0 full-pass rate (GUI + CLI).
Lab claim: WindowsAgentArena 56.2, MacOSWorld 43.0 and MacAgentBench 59.9.
Lab claim: average time per task of 8 min on OSWorld Verified, 1 h 7 min on OSWorld 2.0, 34 min on WeaveBench and 52 min on Agents' Last Exam.
Sources
PAPERCUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements (arXiv)arxiv.org/abs/2611.03126 · Nov 2, 2026LABFirst author Xinlei Yu on Xx.com/vajkcgvd/status/2117559546852114763 · Nov 3, 2026LABCUA-Master project pagexlyu0106.github.io/CUA-Master/REPOxlyu0106/CUA-Master on GitHubgithub.com/xlyu0106/CUA-MasterREPOyxl66666/CUAMaster-27B on Hugging Facehuggingface.co/yxl66666/CUAMaster-27BTracking continues until an independent board lists it.