Feb 8, 2025 at 12:05 PM Hugging Face releases open source rival to challenge OpenAI's Deep Research agent
Less than a day after OpenAI unveiled Deep Research, its internet-based multi-step research tool, Hugging Face has launched Open Deep Research, an open-source AI research agent to reproduce OpenAI's Deep Research results. Developed by Hugging Face's in-house team, this project includes an AI model, OpenAI’s o1, and an agentic framework to assist in planning and tool usage. Although o1 is proprietary and requires a paid API, it reportedly outperforms open models like DeepSeek’s R1.
In just under 24 hours, Hugging Face's researchers adapted o1 to utilize a text-based browser and a “text inspector” toolkit, enabling autonomous web navigation, file manipulation, and data calculations. Despite this rapid development, Open Deep Research scored 55.15% on the GAIA validation sets, trailing behind OpenAI's Deep Research, which achieved 67.36%. Notably, Hugging Face observed optimal performance with a coding agent, whereas using a standard JSON-based agent reduced the score to only 33%.
Hugging Face speculates that OpenAI's superior performance may be attributed to Operator, an agent capable of human-like web interactions. In response, Hugging Face aims to develop its own web browser agent to enhance its capabilities.
- AI Image Generator
- Freemium • Open Source
Related news
External links
- Open-source DeepResearch – Freeing our search agentsHugging Face • Official source
- Hugging Face clones OpenAI’s Deep Research in 24 hoursArs Technica
- Hugging Face researchers aim to build an 'open' version of OpenAI's deep research toolTechCrunch
- Hugging Face makes big progress on open-source Deep Research rivalNeowin
- Don't want to pay for ChatGPT Deep Research? Try this free open-source alternativeZDNET
The whole thing may seem complicated but basically, Hugging Face has connected the open-source agent (i.e. a tool that split complex prompt query into small queries, aka spliting a sentence into the ideas it composes) Magentic-One (Microsoft), to OpenAI's API (it could have been R1 or any other model) to try to match OpenAI's Operator performances. Agents gives much better performances than chatbots alone because latter are unable to split complex queries. So Hugging Face hasn't developed into new tool, just added the glue between the two existing tools, and it gave pretty much OpenAI's Operator, only a little worse. This proves that OpenAI doesn't possess any "magical formula" about AI but has just been able to create a good enough agent and a decent web crawler, with of course an unlimited access to their API. So just as R1 is spreading everywhere, open-source agents (like based on Magentic-One + R1) could be legion very soon. Of course, the final performances depend a lot from the training data used, not only from tools, and since no model disclosure their data sources, tool benchmarks are irrelevant.