4B agent Mage-VL-Cowork scores 66.0% on OSWorld-Verified, beating a 30B model

The Mage team has introduced Mage-VL-Cowork, a computer-use agent with 4 billion parameters. It sees the desktop through screenshots, moves the mouse and presses keys. When needed, it also runs code and terminal commands within the same episode. On OSWorld-Verified, a benchmark of tasks in real desktop applications, it scored 66.0% with a 50-step limit. That beats Muse Glimmer-30B (65.9%), Qwen3.6-27B (63.9%) and EvoCUA-32B (56.7%). Human-level performance on this test is 72.4%.

One agent, two control channels

A graphical interface is universal but expensive: a single screenshot can cost thousands of tokens, and an action that takes one command turns into dozens of clicks. A skilled computer user opens the terminal when it is faster and checks the result in a file, not in the text of a dialog. Mage-VL-Cowork picks the channel itself, within a single policy. This sets it apart from systems like CoAct-1, where an orchestrator splits work between a separate GUI agent and a coder agent.

Access to tools does not guarantee the ability to use them well. Even strong models with a standard setup choose channels poorly. They click where a single library call would do, or edit a file on disk even though the running application holds its state in memory and will overwrite the file on exit. If you distill such trajectories, the student inherits the same habits.

Evolving the harness instead of distilling directly

The authors do not copy the teacher's default behavior. Instead, they evolve its harness: the scaffold, tool interfaces, prompts and terminal environment. An author agent proposes variants and picks the configuration with the best balance of success and cost on debugging tasks. Evolution also includes removal: overlapping or unsuitable tools are dropped so that the teacher produces clean hybrid trajectories.

The result is a set of specialized configurations: a GUI harness, a terminal harness and a hybrid one. The same tasks are run through several of them, which yields trajectories with different emphases.

A task factory and five-stage training

This kind of training needs thousands of diverse environments. The Mage-CUA-Forge pipeline does not build each task from scratch with an open-ended agent. Instead, it creates a reusable workflow for each atomic skill.

The workflow combines LLM nodes with application preparers that set up input artifacts, such as a Calc spreadsheet or a Thunderbird profile with a populated mailbox. The output is a package containing an instruction, an environment setup script and a grading rubric, and verification nodes check consistency before a task is accepted. This process synthesized 293 806 tasks across 6 915 skills, and recurring per-task costs fell by 82.

6%. For training, the team collected 318 837 teacher trajectories, filtered by rubric using an LLM judge and by quality.

Training starts from a general multimodally aligned checkpoint, not from a model that already handles GUI, code and agentic scenarios. There are five stages: GUI mid-training, SFT on GUI agents, SFT on code and terminal, hybrid SFT and online RL. Quality rises step by step: GUI SFT reaches 45.56%, hybrid SFT 55.84% and the final RL stage 66.0%, a gain of 20.44 percentage points. The student's results also guide how the skill taxonomy is expanded and refined between generation rounds.

What this means

A 4B-parameter model achieved the best result among open agents of its size and outperformed models 7–8 times larger. The authors attribute this to the quality of the teacher's harness and the cleanliness of the training data, not to parameter count. For local use cases, this is an argument for compact agents that can run on the user's device at a lower reasoning cost.