Source: Microsoft official model release and technical documentation published October 30, 2026, on Hugging Face, titled 'microsoft/Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use from Evolved Harnesses'.
Built upon the llava_onevision2 architecture, the model fuses Microsoft's custom Mage-ViT visual backbone with a Qwen3-4B language core (Qwen3-4B-Instruct-2507). To resolve the channel arbitration tax that plagues multi-agent systems, the architecture internalizes routing decisions within a three-part turn format: chain-of-thought rationale enclosed in thought tags, an action summary, and a formal tool call. Visual grounding operates across 11 normalized GUI coordinates between 0 and 999, while programmatic operations execute inside an interactive shell environment pre-configured with Python libraries including python-pptx, openpyxl, pandas, and pypdf.
The training curriculum relied on Mage-CUA-Forge, an environment synthesis engine that generated 293,806 distinct tasks spanning 6,915 atomic skills across desktop applications, synthetic InfiniteWeb pages, and SQLite-backed WebHarbor containers. By optimizing the teacher execution harness before trajectory collection—pruning redundant utilities such as xdotool and verbose accessibility trees—engineers lifted the GPT-5.4 teacher to 92.0% accuracy with 29.2% fewer interaction steps, translating into a 12.5 percentage point boost in distilled student capability.
Across task categories in OSWorld-Verified, Mage-VL-Cowork demonstrated marked domain strengths, scoring 87.5% on core Operating System tasks and 62.7% on Multi-Application Workflows. Training consumed 7.79 ZFLOPs of nominal compute across 32 NVIDIA B200 accelerators paired with 1,120 CPU sandbox slots for asynchronous reinforcement learning rollouts. Hardware requirements remain modest, requiring 9.48 GB of bfloat16 memory across two safetensors shards for standard local inference.