A group of developers at AI dev platform Hugging Face, including Thomas Wolf, the company’s co-founder and chief scientist, say they’ve built an “open” version of OpenAI’s deep research tool.
Deep research, which OpenAI unveiled during an event Sunday, crawls the web to compile research reports on any subject. While impressive, deep research is currently only available in limited preview to users subscribed to OpenAI’s $200-a-month ChatGPT Pro plan.
The Hugging Face team’s project, which they’re calling Open Deep Research, consists of an AI model — OpenAI’s o1 — and an open source “agentic framework” that helps the model plan its analysis and guides it to use tools like search engines. O1 is a proprietary model (i.e. gated behind a paid API), but the team says it delivered better performance than “open” models such as DeepSeek’s R1.
In less than 24 hours, the researchers were able to harness o1 to use a simple, text-based browser and a “text inspector” toolkit to read files across the web. Open Deep Research can navigate the web autonomously, the team says, scrolling through pages, manipulating files, and even running calculations with data.
On GAIA, a benchmark for general AI assistants, Open Deep Research achieves a score of 54%. That’s compared with OpenAI deep research’s score of 67.36%.
I tried Open Deep Research in the public demo the team set up — but couldn’t get it to work. The page was under heavy load at publication time; after 10 minutes, it spit out an error message.
But the researchers say that they’re committed to improving the experience, and have made the source code available on GitHub for inspection and feedback.
Worth noting is that there are a number of OpenAI deep research “reproductions” on the web, some of which rely on open models and tooling. The crucial component they — and Open Deep Research — lack is o3, the model underpinning deep research.
Few, if any, models beat o3 on benchmarks related to answering complex questions and information gathering. Short of an open model to rival o3, deep research alternatives may not quite measure up to the real thing.
Source link
