# CUA-Master trains a 27B computer-use agent to 80.2 on OSWorld

> A new open model learns from tutorial videos and its own failures. It trails the best closed models, but it is small and the code and weights are out.

URL: https://claudegpt.si/news/cua-master-computer-use-agent  
Published: 2026-11-04T02:51:11.196Z  
Author: ClaudeGPT Si Newsroom  
Section: Research  
Topics: Research papers, AI agents, Reinforcement learning, Synthetic data, Benchmarks, Qwen  
Publisher: ClaudeGPT Si

## Key takeaways

- CUA-Master-27B, built on Qwen-3.5-27B, scores 80.2 on OSWorld Verified and 36.9 on OSWorld 2.0, according to its authors.
- It trains on about 127,000 samples built from instructional videos and chained tasks, and retries failed steps from saved snapshots.
- It trails closed models such as Claude-Opus-5 (86.3 on OSWorld Verified) and needs heavy infrastructure to reproduce.

CUA-Master is a 27-billion-parameter computer-use agent that scores 80.2 on OSWorld Verified, according to its authors, a paper that Hugging Face listed among its daily papers on November 3, 2026. The team from CUHK, THU, HKUST (Guangzhou) and JD's Explore Academy released the model and the code. The [paper](https://arxiv.org/abs/2611.03126) is mostly about where the training data comes from, and that is the part worth understanding.

A computer-use agent is a model that looks at a screen and operates software the way a person would: it clicks, types, scrolls and, in this case, also runs shell commands. Training one needs thousands of recorded tasks that were carried out correctly. Those are expensive to collect by hand, so most teams generate them with models. The authors say that this approach has three problems: the tasks look alike, the builders have little control over difficulty, and the agent fails more often as tasks get longer, so fewer usable recordings come out.

## How the training loop works

CUA-Master is built on Qwen-3.5-27B and trained in rounds. Each round has two parts.

The first part pulls tasks from outside the model. The team collected instructional videos from YouTube, TikTok and Bilibili, transcribed them with Whisper, and had an annotator model rate how suitable each one was for an agent. Suitable videos were cut into separate tasks, and the steps were rebuilt from the spoken verbs and from changes between video frames.

The second part builds harder tasks from small reusable pieces. The paper calls them primitives: one action in one application, such as changing a setting, with its starting conditions and a checker that confirms it worked. The system chains several primitives, across several applications, into one longer task. The authors control how many primitives (K) and how many applications (N) go into each task.

Between rounds, the team looks at where the current model fails and steers the next round of data toward those gaps. In total the data covers about 127,000 samples across 8 domains and more than 650 applications: 115,000 for supervised fine-tuning and 12,000 for reinforcement learning, the final stage in which the agent practices in real virtual machines and is rewarded when a task's checker passes.

## Snapshot and rollback

Long tasks usually fail somewhere in the middle. Instead of discarding the whole attempt, a detector finds the earliest step where the run went wrong, the environment is restored to a saved snapshot just before it, and the agent tries again with a hint about what went wrong, without being shown the answer. Only the successful branch is kept for training.

The paper's table shows what this buys. For a task chaining 20 primitives across 3 applications, the share of successful runs rises from 19.3% to 56.9%. For 50 primitives across 8 applications, it rises from 0.0% to 18.7%. These are the authors' own measurements on their own generated tasks.

## The results, and what they do not show

The authors report these scores for CUA-Master-27B:

| Benchmark | CUA-Master-27B | Claude-Opus-5 | GPT-5.6-Sol |
| --- | --- | --- | --- |
| OSWorld Verified | 80.2 | 86.3 | 83.6 |
| OSWorld 2.0 | 36.9 | 64.2 | 59.0 |
| WeaveBench | 43.9 | 55.7 | 50.6 |
| Agents' Last Exam | 25.0 | 28.3 | 26.3 |

Against a comparable 27B specialist model, UI-Mate-27B, it gains 2.6, 12.9, 25.6 and 11.8 points on the four benchmarks. On WeaveBench, which mixes graphical and command-line work, it is about level with the closed GPT-5.6-Luna (43.9 versus 44.0) and runs about 28% faster.

It does not beat the best models. The closed models in the paper's table lead on all four benchmarks, and the 2.8-trillion-parameter open model Kimi-K3 reaches 55.9 on OSWorld 2.0. On OSWorld 2.0, CUA-Master is barely ahead of the stock Qwen-3.8-27B (36.5). The paper calls the OSWorld Verified result competitive, which is fair, but it is not a lead.

The other finding is about tools. On WeaveBench, an agent limited to the graphical interface scored 6.6 and one limited to the command line scored 11.5. The hybrid agent scored 43.9 with 23.8% fewer steps than the graphical-only one.

## Limits

The authors name two. The method needs a lot of infrastructure: the paper describes 128 NVIDIA H200 GPUs for training and an environment layer that supports over 2,000 concurrent virtual machines. And the generated data depends on a strong annotator model, which the paper does not name. Stronger annotators gave larger gains, so a team without access to one may get less.

Two more cautions. All the numbers come from the paper, and we have not seen independent reproductions. The text we read also does not state a license for the code or the weights, so check the repositories before building on them.

## Why it matters

The pattern is the useful part. Rather than letting a model train on its own output until it plateaus, the team keeps feeding in material from outside, such as human tutorial videos, and aims it at what the model gets wrong. That makes a 27B model usable on tasks that previously needed far larger ones, though not yet the best of them. Watch the independent OSWorld leaderboard to see whether the numbers hold up outside the paper. Our earlier piece on [training tool-use agents](https://claudegpt.si/news/trajectory2gym-tool-use-agents-training) covers a related approach to agent training data.

Links: [paper](https://arxiv.org/abs/2611.03126), [code](https://github.com/xlyu0106/CUA-Master), [model weights](https://huggingface.co/yxl66666/CUAMaster-27B), [project page](https://xlyu0106.github.io/CUA-Master).

## Frequently asked questions

### What is a computer-use agent?

It is an AI model that looks at a screen and operates software like a person: clicking, typing and scrolling. CUA-Master also runs shell commands.

### Is CUA-Master open source?

The authors released the code on GitHub and the 27B weights on Hugging Face. The paper text does not state a license for either, so check the repositories before using them.

### What is OSWorld Verified?

It is a benchmark of real computer tasks that an agent must complete in a virtual desktop. CUA-Master-27B scores 80.2 on it, according to its authors.

## Sources

- [CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements](https://arxiv.org/abs/2611.03126), arXiv
- [CUA-Master on Hugging Face Papers](https://huggingface.co/papers/2611.03126), Hugging Face
- [CUA-Master repository](https://github.com/xlyu0106/CUA-Master), GitHub
- [CUAMaster-27B](https://huggingface.co/yxl66666/CUAMaster-27B), Hugging Face
- [CUA-Master project page](https://xlyu0106.github.io/CUA-Master), CUA-Master authors