Microsoft open-sources Mage-VL-Cowork: 4B hybrid GUI+terminal computer-use agent hits 66.0% on OSWorld-Verified

Why it matters

Microsoft’s Mage team open-sourced Mage-VL-Cowork (Apache-2.0) on Hugging Face on 2026-10-30: a 4B hybrid computer-use agent that combines screenshot-grounded GUI actions with persistent bash/code execution in a single VLM policy. It reports 66.0% on OSWorld-Verified (361 tasks, 50-step budget, two-run average), best at 4B, with 87.5% on OS tasks and 62.7% on multi-app workflows (vs GPT-5.4-mini 69.6%, ScaleCUA-Qwen3.5-9B 68.7%, Muse Glimmer-30B 65.9%; human 72.4%). Built on Mage-ViT + Qwen3-4B, trained via a five-stage specialize-then-unify curriculum ending in online GRPO (+10.16 pp over hybrid SFT, fewer calls/tokens), ~7.79 ZFLOPs on 32 B200s. Evolving the GPT-5.4 teacher harness lifted the teacher to 92.0% and the student by +12.5 pp; the Mage-CUA-Forge synthesis pipeline cut per-task cost from $0.906 to $0.158.