Mage VL Cowork

Release Summary

Microsoft's open-weight 4B computer-use model. It acts through screenshots and a persistent terminal in one checkpoint, with no external orchestrator. Apache 2.0 weights at microsoft/Mage-VL-Cowork. Image and text in, text out. The model card documents a 32,768-token inference window. The architecture supports up to 262,144 tokens, and that longer ceiling is not the published inference window, so this catalog uses 32,768 for both context and output. On OSWorld-Verified, with the MAGEHybridAgent harness, a 50-step budget, and 361 tasks excluding Google Drive, the two-run average is 66.0% (65.33% and 66.75%). Category scores on that table are 87.5% for OS, 64.9% for Office, 65.0% for Daily, 66.3% for Professional, and 62.7% for multi-application workflows. Muse Glimmer 30B is 65.9% on the same table. The released checkpoint uses 18.77 model calls and 5,499 output tokens per task on that evaluation. No hosted token price and no measured price per task.

Timeline