Microsoft open-sources innovation framework: DeepSeek-R1 and other models can be turned into AI Agents

Odaily report: Microsoft has released the latest version 2.0 of its visual Agent parsing framework OmniParser on its official website, which can turn models like DeepSeek-R1, GPT-4o, and Qwen-2.5VL into usable AI Agents on computers. Compared to V1, V2 has higher accuracy and faster inference speed when detecting smaller interactive UI elements, with a 60% reduction in latency. In the high-resolution Agent benchmark ScreenSpot Pro, the accuracy of V2+GPT-4o reached an impressive 39.6%, while the original accuracy of GPT-4o was only 0.8%, a significant overall improvement. In addition to V2, Microsoft has also open-sourced omnitool, a Docker-based Windows system that covers screen understanding, positioning, action planning, and execution, which is a key tool for turning large models into Agents.