In another remarkable demonstration of the incredible speed at which the AI market is moving, a team at Hugging Face has replicated Deep Research, the OpenAI agent product which was released earlier this week.
The team’s new Open DeepResearch is an open source initiative which has already matched the benchmark performance of the OpenAI product, at a fraction of the cost.
Deep Research is an AI agent which can conduct complex multi-step web research using reasoning and a base LLM, in this case the unreleased o3 model. OpenAI claims that its new agent is clever enough to pivot mid-way and point in a new direction in its research if new data makes it necessary.
The company states: “Powered by a version of the upcoming OpenAI o3 model that’s optimized for web browsing and data analysis, it leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet, pivoting as needed...”
Open DeepResearch came just 24 hours after the launch of the commercial alternative, with a goal to provide similar functionality, but for free and open so it can be used by anyone around the world.
There are several major challenges involved in this attempt, as the open agent must combine multimodal support with reasoning and task chaining, to ensure that it can effectively project-solve on the fly.
Open DeepResearch came just 24 hours after the launch of the commercial alternative, with a goal to provide similar functionality, but for free.
An important target for the team is to approach parity with Deep Research on the GAIA leaderboard, a table that validates autonomous agent frameworks. OpenAI reached 67.36% on the validation benchmark, compared to just 9.7% from the bare GPT-4 Turbo LLM without agentic support.
The Hugging Face team to date has totalled a very credible 55.15% average. These scores reflect an agent’s ability to autonomously cope with 450 non-trivial questions and solve them.
The team posted: “In our 24h+ reproduction sprint, we’ve already seen steady improvements in the performance of our agent on GAIA! We’ve quickly gone up from the previous SoTA...to our current performance of 55.15% on the validation set.”
The Hugging Face initiative is a work in progress. The developers are inviting the public to contribute to help build out the final product. For example, the project website details the need to produce an improved multimodal web browser, additional agents and local model support.
Those who are interested in seeing what progress has been made so far, should visit the demo site directly. This is a Hugging Face Space, but it appears to be extremely overloaded at the moment. When I tried to run an agent I received an error message, which I’m assuming reflects teething problems.
Still this is definitely a project worth watching.
More from Tom's Guide
- ChatGPT's new free o3-mini model is mind-blowing — 5 prompts to try first
- I tested ChatGPT o3-mini vs DeepSeek R1 vs Qwen 2.5 with 7 prompts — here’s the winner
- Verizon offering epic deal on Google One AI Premium for 50% off
