Microsoft releases Mage-VL-Cowork 4B hybrid computer-use model

The training pipeline, called Mage‑CUA‑Forge, generated 293,806 tasks covering 6,915 atomic skills, cutting recurring per‑task generation cost by 82.6% (from $0.906 to $0.158) and token usage by 93.3% (1,276 K → 85 K tokens). Online hybrid reinforcement learning added +10.16 pp accuracy while reducing mean model calls by 13.0% and output tokens by 25.4%. The model is released with a Docker container and SGLang server support, using a custom multimodal architecture (llavaonevision2) and a 32,768‑token context window.