Microsoft releases 4B model Mage-VL-Cowork, which solves 66.0% of OSWorld-Verified tasks

Microsoft has published Mage-VL-Cowork on Hugging Face, an open 4B-parameter model for desktop agents (CUA, where an AI works on a computer by itself). On the OSWorld-Verified benchmark, where agents solve tasks in real operating systems, it handled 66.0% of 361 held-out tasks with a budget of 50 steps. Until now, two approaches have coexisted in this field: one agent sees only screenshots and clicks with the mouse, while the other writes code and edits files directly. Microsoft combined both in one compact model, and the model itself chooses which path to take at each step.

Two action channels inside one policy

The system has no external router. The model decides for itself which tool it needs at any moment. For visual layout, canvas edits and work with the live state of an application, it uses the computer_use channel: 11 GUI primitives and normalized coordinates. For structured edits in spreadsheets and documents, and for checking results after the work is done, it switches to bash, a persistent shell with a Python environment.

The logic is straightforward. Changing an xlsx file through dozens of menu clicks is slower and more fragile than writing the file with openpyxl. On the other hand, a script cannot fix a visual effect in an image.

The 66.0% score on OSWorld-Verified is averaged over two full runs on 361 held-out tasks. In the same table, Qwen3.6-27B scored 63.9% and Muse Glimmer-30B scored 65.9%. ScaleCUA-Qwen3.5-9B, however, reached 68.7%, which is higher. So it is not accurate to say the small model beats all its competitors.

How Microsoft built the training data

The first observation concerns the teacher. A strong model given an excess of tools picks up bad habits, and it passes them on to the student during distillation. Microsoft removed redundant channels such as xdotool and bulky accessibility trees, and preinstalled the necessary Python libraries. As a result, the GPT-5.4 teacher improved to 92.0% while using 29.2% fewer steps. The student trained on those trajectories scored 37.6% versus 25.1%, a gain of 12.5 percentage points.

The second is task generation. The Mage-CUA-Forge pipeline writes a generating workflow once for each atomic skill, rather than running an open-ended agent loop for each task. This produced 293 806 tasks across 6 915 skills. The cost per task fell from 0.906 to 0.158 dollars, a drop of 82.6%.

The third is the five training stages. First, visual grounding to elements (grounding), GUI execution and terminal work are trained separately. Then Hybrid SFT teaches the model to switch between channels. The final stage is online GRPO with a reward for a verifiable task outcome. Quality rises from 45.56% to 55.84% and then to 66.0%. Online RL added 10.16 percentage points and at the same time cut the average number of model calls from 21.57 to 18.77, and output tokens from 7 374 to 5 499.

The context is organized asymmetrically. It retains the three latest screenshots and four full responses, while earlier history is compressed into one-line Action summaries. The active window is 32K tokens. The bf16 weights take up about 9.48 GB, and the license is Apache-2.0.

What teams should keep in mind

This matters most to those building enterprise assistants. If your agent by default imitates a human and clicks on every interface element, you now have a downloadable, reproducible model to compare against. Evaluation changes too: looking at the share of solved tasks is not enough, and teams also need to check whether the chosen path made sense and whether the final file is correct.

The limitations are significant as well. The 66.0% figure is claimed by the vendor itself, and the same table contains a stronger 9B model. Fewer calls and tokens do not guarantee that users will wait less: a short script may be paired with a slow application launch. Finally, the five-stage training cost a claimed 7.79 ZFLOPs on 32 B200 GPUs. A small model saves on inference, but not on training costs. Open weights and the recipe allow independent teams to reproduce the result, and their reruns will show whether 66.0% holds up under identical conditions.