Microsoft published Mage-VL-Cowork on Hugging Face on 30 October under Apache-2.0. It is a 4B vision-language model for computer use, built from Microsoft’s Mage-ViT image encoder and Qwen3-4B-Instruct-2507. What sets it apart is one policy that decides at every turn whether to act on the screen with mouse and keyboard or to drop to a bash command.
The numbers, and how they were measured
- OSWorld-Verified: 66.0%, averaged over two runs (65.33 and 66.75), on 361 tasks with a 50-step budget. The 361 tasks leave out the Google Drive set, so this is not the full benchmark.
- In Microsoft’s own table: it is ahead of Qwen3.6-27B (63.9) and Muse Glimmer-30B (65.9), but behind ScaleCUA-Qwen3.5-9B (68.7), and well behind frontier proprietary models listed at 75 to 78.7. The human reference is 72.4.
- WebVoyager-Static: 58.6%.
- Size: about 9.5 GB in bf16 with a 32,768-token context, so it fits on a 12 GB or 16 GB card.
How it was trained
The project page describes Mage-CUA-Forge, a pipeline that produced 293,806 training tasks and cut the cost per task from $0.906 to $0.158. Microsoft says pruning and evolving the teacher’s tool setup before collecting examples lifted the 4B student by 12.5 points on 672 held-out multi-application tasks (37.6% against 25.1%), with the GPT-5.4 teacher at 92.0% on the same tasks.
Why this matters
Computer-use agents have mostly needed a frontier API, which means sending screenshots of your desktop to someone else’s servers. A 4B model that runs locally and mixes GUI and shell actions makes private desktop automation practical on a single GPU, compared with paid options like GPT-6.1 Terra. Treat the scores as a starting point and run it in a sandboxed VM, since an agent with bash access can do real damage when it guesses wrong.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment