Open-source DeepResearch – Freeing our search agents

Hugging Face · 1 year ago · 21 ● 2 sources

A team released an open-source reproduction of OpenAI's Deep Research system, which uses an AI agent to browse the web and answer complex questions. Their implementation achieved 55.15% accuracy on the GAIA benchmark's validation set, up from 46% with the previous open-source framework, primarily by having the agent write actions in code rather than JSON. The open-source version enables anyone to run a web-searching AI agent locally with their choice of models, though the team plans further improvements including vision-based web browsing and GUI interaction capabilities.