VibeusMicrosoft’s Mage-VL-Cowork combines screenshot-based desktop... - Vibeus
Microsoft’s Mage-VL-Cowork combines screenshot-based desktop interaction with persistent code and terminal execution in one compact vision-language policy. It chooses when to click and when to edit files programmatically, without an external agent router. Microsoft describes it as a 4B model; the BF16 checkpoint is 9.48 GB.
The team reports 66.0% on OSWorld-Verified: 361 held-out tasks excluding Google Drive, a 50-step budget, and an average of two full runs, 65.33% and 66.75%. OS tasks reached 87.5%; multi-application workflows 62.7%. These are the team’s benchmark results.
Training first develops visual grounding, GUI execution and terminal/coding skills, then combines them through hybrid SFT and online RL. Mage-CUA-Forge generated 293,806 tasks across 6,915 skills. Compared with hybrid SFT, online RL added 10.16 percentage points while reducing model calls 13.0% and output tokens 25.4%.
The active context window is 32,768 tokens, retaining three recent screenshots and four full responses plus compressed action history. The release includes reference agent code and evaluation instructions under Apache-2.0.
It uses a custom architecture and requires an execution harness and isolated sandbox; the weights alone are not a ready desktop app. The official deployment guide calls for human confirmation before irreversible actions.
