{"version":1,"type":"story","url":"https://digestai.news/story/microsoft-releases-mage-vl-cowork-4b-hybrid-computer-use-model","json":"https://digestai.news/story/microsoft-releases-mage-vl-cowork-4b-hybrid-computer-use-model.json","markdown":"https://digestai.news/story/microsoft-releases-mage-vl-cowork-4b-hybrid-computer-use-model.md","slug":"microsoft-releases-mage-vl-cowork-4b-hybrid-computer-use-model","headline":"Microsoft releases Mage-VL-Cowork 4B hybrid computer-use model","summary":"Microsoft unveiled Mage‑VL‑Cowork, a 4 billion‑parameter vision‑language model that can both interpret GUI screenshots and execute terminal commands without an external orchestrator. Trained from an intermediate Mage‑VL checkpoint through a five‑stage specialize‑then‑unify curriculum, the model reaches 66.0% success on the OSWorld‑Verified benchmark, the highest score among 4 B models, and scores 87.5% on pure OS tasks and 62.7% on multi‑application workflows.\n\nThe 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.","keyPoints":["Mage‑VL‑Cowork achieves 66.0% OSWorld‑Verified score, best among 4 B models","Training pipeline generated 293,806 tasks, lowering per‑task cost to $0.158 (‑82.6%)","Online hybrid RL improves accuracy by +10.16 pp and cuts token usage by 25.4%"],"whyItMatters":"A compact 4 B multimodal model that reliably handles full‑desktop and terminal tasks narrows the gap between large‑scale agents and on‑device AI, enabling more efficient, privacy‑preserving automation.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["Microsoft","Hugging Face"],"models":["Mage-VL-Cowork","GPT-5.4","Qwen3.5-4B","Qwen3-VL-4B","Fara1.5-4B","GPT-5.4-mini"],"people":[]},"firstPublishedAt":"2026-10-31T10:02:47Z","updatedAt":"2026-10-31T10:02:47Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"huggingface.co","title":"microsoft/Mage-VL-Cowork · Hugging Face","url":"https://huggingface.co/microsoft/Mage-VL-Cowork","publishedAt":"2026-10-31T10:02:47Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1xk6blw/microsoftmagevlcowork_hugging_face/","points":null}],"thread":null,"cite":{"text":"Digest AI, \"Microsoft releases Mage-VL-Cowork 4B hybrid computer-use model\", 31 October 2026, https://digestai.news/story/microsoft-releases-mage-vl-cowork-4b-hybrid-computer-use-model","publisher":"Digest AI","title":"Microsoft releases Mage-VL-Cowork 4B hybrid computer-use model","datePublished":"2026-10-31T10:02:47Z","url":"https://digestai.news/story/microsoft-releases-mage-vl-cowork-4b-hybrid-computer-use-model"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}