Native computer use offers a general interface for agents to operate almost any software available to people, but requires long-horizon state tracking, large-scale interactive experience, and learning from sparse yet verifiable outcomes. We introduce Qwen-CUA, a native computer-use agent with a 397B-A17B Qwen mixture-of-experts backbone. It observes only screenshots and acts through keyboard and mouse events, without DOM trees, accessibility metadata, or task-specific APIs. Its scaffold maintains up to 20 active screenshots and folds older visual history in fixed-size blocks to retain recent evidence while preserving reusable prompt prefixes. For training, we build a cloud rollout fleet with access to nearly 100,000 vCPUs and tens of thousands of concurrent environments, construct approximately 40,000 verifiable tasks, and collect personalized long-horizon workflows across everyday and professional software. We optimize complete trajectories with verifiable rewards and trajectory slicing, while iterative training runs refresh supervised data and recalibrate reinforcement-learning tasks. Across eight benchmarks, Qwen-CUA outperforms Qwen3.7 and remains competitive with leading proprietary systems, reaching 86.2 on OSWorld-Verified and 18.5/48.4 binary/partial completion on OSWorld 2.0. Scaling the same recipe to a model with over one trillion parameters yields Qwen-CUA-Max, improving these scores to 87.6 and 21.2/53.3. Qwen-CUA also reduces RedTeamCUA attack success from 36.6 to 16.4 relative to Qwen3.7. Efficiency analyses, a browser deployment, and Bash-augmented experiments further characterize practical behavior. These results establish native computer use as a broadly capable agent foundation and highlight scalable verifiable interaction and hybrid tool use as key directions.
核心贡献 · Key contributions
提出 Qwen-CUA，一个基于 397B-A17B MoE 架构的原生计算机使用智能体，仅通过截图和键盘/鼠标事件操作。 Introduces Qwen-CUA, a native computer-use agent with a 397B-A17B MoE backbone, operating solely via screenshots and keyboard/mouse events.
利用约 10 万 vCPU、4 万任务以及轨迹切片和 SAPO 的迭代强化学习，扩展可验证的交互训练。 Scales verifiable interactive training with ~100k vCPUs, 40k tasks, and iterative RL with trajectory slicing and SAPO.
在八个基准上取得领先结果，包括 OSWorld-Verified 86.2 和 OSWorld 2.0 18.5/48.4。 Achieves state-of-the-art results on eight benchmarks, including 86.2 on OSWorld-Verified and 18.5/48.4 on OSWorld 2.0.
展示鲁棒性提升，将 RedTeamCUA 攻击成功率从 36.6 降至 16.4。 Demonstrates robustness improvements, reducing RedTeamCUA attack success from 36.6 to 16.4.
展示扩展到万亿参数的 Qwen-CUA-Max 进一步提升性能，达到 87.6 和 21.2/53.3。 Shows scaling to trillion-parameter Qwen-CUA-Max further improves performance, reaching 87.6 and 21.2/53.3.
强调与 Bash 的混合工具使用，缩短轨迹，并推动原生计算机使用作为通用基础层。 Highlights hybrid tool use with Bash, shortening trajectories and motivating native computer use as a universal grounding layer.
局限 · Limitations
基于像素的观察和串行低级动作引入延迟和交互成本，限制效率。 Pixel-based observation and serialized low-level actions introduce latency and interaction cost, limiting efficiency.
对于可通过代码或 API 表达的任务，视觉操作效率低下，混合路由仍是开放挑战。 Visual operation is inefficient for tasks expressible via code or APIs, requiring hybrid routing that remains an open challenge.
在 RedTeamCUA 上仍有 16.4%的攻击成功率，存在残余安全风险，并非部署安全保证。 Residual security risk persists with 16.4% attack success on RedTeamCUA, not a deployment-safety guarantee.
基于回合的效率比较受串行执行影响，缺乏标准化的低级动作度量。 Turn-based efficiency comparisons are confounded by serialized execution and lack normalized low-level action measures.
训练依赖昂贵的基础设施和可验证任务构建，可能限制可复现性和可扩展性。 Training relies on costly infrastructure and verifiable task construction, which may limit reproducibility and scalability.
论文章节 · Sections（共 21）
Qwen-CUA：面向（几乎）一切任务的原生计算机使用Qwen-CUA: Native Computer Use for (almost) Everything