A group of developers holding the AI DEV platform face, including the company's co -founder and chief scientist THOMAS WOLF, has built an open version of Openai's Deep Research Tool.
Deep Research announced by Openai during Sunday events is a CRA to edit research reports on all subjects. Impressive, but now, deep -search is only available in previews limited to users who have subscribed to Openai's $ 200 Chatgpt Pro Plan.
The hug face team project that calls the open deep -search is the “Agent Framework”, which is an AI model (O1 of OPENAI), a model planning to use tools such as search engines. Is composed of. O1 is a unique model (that is, the gate behind a paid API), but the team has provided better performance than “Open” models such as DeepSeek R1.
In less than 24 hours, researchers were able to read the files throughout the web using the O1 using the text -based browser and the “Text Inspector” toolkit. Open deep -search states that you can autonomously navigate the web by scrolling pages, operating files, executing data using data.
GAIA, a general AI assistant benchmark, has achieved 54 % of the open deep -search scores. This is compared to the 67.36 % score of Openai Deep Research.
I opened a deep -searching demonstration set by the team, but I couldn't work. This page had a heavy load at the time of publication. After 10 minutes, spit out the error message.
However, researchers say they are committed to improving the experience, making the source code available on GitHub for testing and feedback.
What is worth noting is that there are a lot of Openai's deep research “duplication” on the web, which depends on open models and tools. Their important elements and the lack of a deep research are O3 and a model that supports deep research.
If so, there are few models that break O3 with a benchmark related to complicated questions and answers to information gathering. In the range of open models comparable to O3, deep research alternatives may not be measured in real one.