Open-source DeepResearch in 24 hours

What happened

OpenAI released Deep Research, a system that browses the web and answers questions from what it reads. Hugging Face’s post the next day describes a 24-hour effort to reproduce it openly.

The technical details

From the post:

• OpenAI reported large gains on GAIA: near 67% correct on average (1-shot), and 47.6% on the hardest “level 3” questions
• Deep Research combines an LLM with an agentic framework that guides tool use (like web search) in steps, and OpenAI disclosed little about that framework
• the open version uses a CodeAgent (from smolagents, the previous post) plus custom browsing and text-inspection tools
• the post includes results, community reproductions and next steps

Why it matters

It shows that much of an agent’s power is in the framework around the model (tools, loop, memory), not only the model. Two months later, “Code as Agent Harness” (May 2026 on this blog) builds a whole survey on that idea.