Many desktop tasks are awkward when an agent must choose between seeing the interface and acting on structured files. The key insight here is to make that choice within one policy: use the GUI when visual state matters, and use code when a file or system operation is more direct and verifiable. In a 361-task OSWorld-Verified evaluation with a 50-step budget, it scored 66.0% at 4B parameters; that result is notable not just for the score, but for the compact model size.
Key Capabilities
- It routes each turn between screenshot-grounded GUI actions and a persistent shell, without relying on an external router. That allows it to combine visual editing with direct spreadsheet, document, or system operations.
- Training follows a five-stage specialize-then-unify curriculum: visual grounding, GUI execution, coding and terminal skills, hybrid supervision, then online reinforcement learning. This gives the model separate competencies before asking it to coordinate them.
- The training pipeline reports 293,806 synthesized tasks across 6,915 skills. Online RL raised OSWorld-Verified performance by 10.16 percentage points over hybrid supervised fine-tuning while reducing average output tokens by 25.4%.
- Its strongest reported OSWorld category is operating-system tasks at 87.5%; it reaches 62.7% on multi-application workflows. These scores are benchmark results, not a guarantee of reliable behavior on arbitrary desktops.
Who It’s For
A good fit for researchers evaluating compact computer-use agents, or teams building controlled desktop automation that benefits from both visual interaction and direct file handling. Look elsewhere if you need a ready-to-use consumer assistant, broad production guarantees, or an agent that can safely run without supervision: its shell and GUI actions can change files and system state, so isolated sandboxes and human approval for consequential actions are important.
How the Training Design Fits Together
The work treats data quality and tool design as part of the model recipe, not as incidental infrastructure. It first simplifies the teacher’s action routes, then uses reusable skill-conditioned workflows to generate tasks and trajectories. This matters because a student can inherit a teacher’s confused tool habits; the reported comparison found that pruning redundant routes improved the 4B student’s success rate by 12.5 percentage points.
