GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
← Back to feed

Qwen-CUA Hits 86.2% on OSWorld-Verified Using Only Screenshots and a Mouse

Most computer-use agents quietly depend on accessibility trees, application APIs, or structured screen representations. Qwen-CUA strips all of that out. The model sees screenshots, clicks the mouse, and types on a keyboard. That is the entire interface.

The paper, arxiv 2608.02352, is titled “Qwen-CUA: Native Computer Use for (almost) Everything.” It was published August 7.

The result

OSWorld-Verified: 86.2%. OSWorld is a benchmark of real desktop tasks across standard applications — file managers, web browsers, productivity software. The Verified subset filters for tasks with reliable, automated grading. At 86.2%, Qwen-CUA posts the highest published result using a screenshots-only interface.

The model has 397 billion parameters. The architecture is not disclosed in detail but the parameter count indicates a sparse MoE configuration. It is not the same model as Qwen3.8-Max.

How it was trained

Training used approximately 100,000 virtual processor cores and around 40,000 checkable tasks spread across diverse software environments. The scale matters because computer-use training requires interactive rollouts: the agent acts, the environment responds, the agent adapts. Unlike text or code, there is no static dataset to distil from. Every training example had to be generated through live interaction.

The reward structure evaluates the final software state, not the path taken to it. Multiple valid action sequences can reach the same correct endpoint, so the model is not penalised for taking a different route than the reference. On long tasks, training sequences are split into manageable pieces to keep gradient signal effective.

The memory design

Qwen-CUA keeps 20 recent screenshots in context. Older frames are folded away. The model also preserves a record of its prior actions, so it can reason about what it has already done on a task that spans many steps. This addresses a key failure mode in computer-use: forgetting mid-task state when the visual context scrolls out of view.

The design is explicitly non-specialist. No website code. No accessibility labels. No task-specific tool wrappers. The only inputs are what a human looking at the screen would see.

Where it still fails

Long-horizon tasks remain the hard case. There is a documented gap between partial task completion and full task completion on multi-step sequences. The model makes progress across the full task distribution — no category of software causes total failure — but the rate at which it fully resolves complex, multi-stage tasks with irreversible intermediate states lags behind its overall OSWorld score.

The difficulty compounds when the task changes state irrecoverably partway through. A wrong click that closes a dialog, an accidental file deletion, an application crash mid-flow — each of these forces the model to detect the state change and adapt, which current screenshot-only agents handle inconsistently.

Why the interface restriction matters

Removing accessibility labels and application APIs has an architectural consequence: the same model runs against any software without modification. There is no per-application tuning, no schema negotiation, no reliance on developers exposing structured interfaces. The screenshot becomes the universal contract.

This approach trades a performance ceiling (accessibility APIs give direct semantic information) for deployment generality. At 86.2% on a standardised benchmark, Qwen-CUA’s results suggest the performance ceiling from dropping structured access is narrower than previously assumed.

The implication for enterprise deployment is practical: a single agent can operate legacy applications, custom internal tools, and standard software without requiring developers to maintain machine-readable interfaces for each.

Key Numbers

MetricValue
OSWorld-Verified86.2%
Parameters397B
Training compute~100,000 virtual CPUs
Training tasks~40,000
Context: recent screenshots20 frames
InterfaceScreenshots, mouse, keyboard only
Paperarxiv 2608.02352
PublishedAugust 7, 2026