Qwen-CUA: Native Computer Use for (almost) Everything
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - Qwen-CUA is a native computer-use agent built on a 397B-A17B Qwen MoE backbone that operates software purely from screenshots plus keyboard/mouse events, with no DOM, accessibility metadata, or task-specific APIs. It shows that large-scale verifiable interactive RL can push open models to near-proprietary performance on general GUI control.
- Scaffold keeps up to 20 active screenshots and folds older visual history into fixed-size blocks, retaining recent evidence while preserving reusable prompt prefixes for caching.
- Training used a cloud rollout fleet with ~100,000 vCPUs and tens of thousands of concurrent environments, ~40,000 verifiable tasks, and personalized long-horizon workflows; optimization used full-trajectory verifiable rewards with trajectory slicing plus iterative SFT/RL refresh cycles.
- Reaches 86.2 on OSWorld-Verified and 18.5/48.4 binary/partial completion on OSWorld 2.0 across eight benchmarks; scaling the same recipe past one trillion parameters (Qwen-CUA-Max) gives 87.6 and 21.2/53.3.
- Safety improves alongside capability: RedTeamCUA attack success drops from 36.6 to 16.4 versus Qwen3.7, and the paper adds efficiency analyses, a browser deployment, and Bash-augmented hybrid tool-use experiments.
Sources (1)
Qwen-CUA: Native Computer Use for (almost) Everything
TL;DR - Qwen-CUA is a native computer-use agent built on a 397B-A17B Qwen MoE backbone that operates software purely from screenshots plus keyboard/mouse events, with no DOM, accessibility metadata, or task-specific APIs. It shows that large-scale verifiable interactive RL can push open models to near-proprietary performance on general GUI control.
- Scaffold keeps up to 20 active screenshots and folds older visual history into fixed-size blocks, retaining recent evidence while preserving reusable prompt prefixes for caching.
- Training used a cloud rollout fleet with ~100,000 vCPUs and tens of thousands of concurrent environments, ~40,000 verifiable tasks, and personalized long-horizon workflows; optimization used full-trajectory verifiable rewards with trajectory slicing plus iterative SFT/RL refresh cycles.
- Reaches 86.2 on OSWorld-Verified and 18.5/48.4 binary/partial completion on OSWorld 2.0 across eight benchmarks; scaling the same recipe past one trillion parameters (Qwen-CUA-Max) gives 87.6 and 21.2/53.3.
- Safety improves alongside capability: RedTeamCUA attack success drops from 36.6 to 16.4 versus Qwen3.7, and the paper adds efficiency analyses, a browser deployment, and Bash-augmented hybrid tool-use experiments.