Microsoft's 4B Mage-VL-Cowork Hits 66% on OSWorld-Verified, Near Human Level
microsoft · hf · 2026-11-04
Microsoft unveils Mage-VL-Cowork, a 4B hybrid computer-use agent trained with a progressive curriculum and scalable synthetic data generation, achieving SOTA 66.0% on OSWorld-Verified.
From the comparison chart:
- 66.0% tops Muse Glimmer-30B (65.9) among on-device models and approaches the human level of 72.4%
- Far ahead of same-size peers like Qwen3.5-4B (35.6) and Fara1.5-4B (13.8)
- Also beats several frontier models: Claude Sonnet 4.5 (62.9), Seed-1.8 (61.9), UI-TARS-2 (53.1), OpenAI CUA (31.3)
The core contribution is a full training recipe learning hybrid computer use from evolved harnesses with efficient reasoning.
More from coding & agent
- Assigning Agents Standard/High/Max Effort Levels Is Just Story Points for Machines — feregri_no · 2026-11-04
- 博主晒海外 AI 数字员工分工：GPT 管后端、Claude 剪片、Gemini 写文案 — lxfater · 2026-11-04
- Aixio 推出托管 MCP 服务，让编码 Agent 直连图像视频设计工具 — aiiiiiaiaiaiaiwater · 2026-11-04
- Opsec runs a fleet of parallel AI security agents that find exploits and auto-open fix PRs — AccBalanced · 2026-11-04
- Teknium clarifies Hermes Agent works with any subscription, cloud or otherwise — Teknium · 2026-11-04
- Alibaba's adaptive GUI agent training cuts environment interactions by 56-61% — Yuxiang Chai · 2026-11-04