OpenAI’s Deep Research, an AI-driven assistant designed to automate complex research tasks, has sparked a rapid response from Hugging Face.
Led by co-founder and chief scientist Thomas Wolf, a small team at Hugging Face has started developing Open DeepResearch, an open-source alternative aiming to provide similar research automation capabilities while remaining freely available.
Deep Research, which OpenAI just integrated into ChatGPT Pro, and up to the name seems to be inspired by Google Gemini’s Deep Research feature, is positioned as a tool for professionals in academia, law, and finance. It autonomously scans online sources, processes data, and generates structured research outputs.
However, being exclusive to ChattGPT Pro users at $200 per month limits access to a limited group of users who are willing to pay such a high price for a subscription.
In response, Hugging Face has released Open DeepResearch as a community-driven project built on publicly available frameworks. The goal is to provide a research assistant that is open, transparent, and customizable, offering an alternative to OpenAI’s increasingly closed AI ecosystem.
How Open DeepResearch Works
Hugging Face’s approach relies on an agent-based AI framework, which structures research tasks into sequential steps. The tool can autonomously browse the web, extract key insights, and generate research summaries.
Unlike OpenAI’s Deep Research, which uses the proprietary o3 model, Open DeepResearch currently operates using OpenAI’s o1 model (or other models available via the OpenAI-API), as fully open-source alternatives have not yet reached comparable performance levels. The team initially tested DeepSeek R1, but early trials showed that it struggled with accuracy and reasoning.
According to Hugging Face, OpenAI’s Deep Research feature seems to integrate with Operator — its new, proprietary AI agent that visually interacts with web content. Open DeepResearch instead uses a text-based web scraper. This means it cannot analyze images, videos, or interactive elements, which limits its ability to handle multimedia-rich sources.
To improve Open DeepResearch’s browsing and text inspection capabilities, Hugging Face has incorporated elements inspired by Microsoft’s Magentic-One, a modular multi-agent AI system designed for complex task automation.
Magentic-One, introduced as part of Microsoft’s AutoGen 0.4 framework, coordinates specialized AI agents for tasks such as web navigation, file analysis, and code execution. While Open DeepResearch does not integrate Magentic-One directly, it adopts a similar agent-based approach to structuring AI workflows.
OpenAI’s Deep Research is designed to autonomously make decisions across multiple sources, leveraging its powerful and so far unreleased o3-model for deeper reasoning.
Hugging Face’s open-source model, running on o1 – based on what OpenAI offers via API, is less capable in this regard, though it still performs structured research queries with reasonable accuracy. (models can be selected from the current list of LLMs provided by OpenAI, 4o, o1, o3-mini, etc.)
Early Performance and Benchmarks
To assess its effectiveness, Hugging Face says it tested Open DeepResearch using the GAIA benchmark, which measures AI performance in autonomous research tasks.
OpenAI’s Deep Research, leveraging o3, scored 67.36%, while Open DeepResearch achieved 54%. While OpenAI’s tool remains ahead, the results suggest that open-source alternatives could narrow the gap with further optimization.
One critical finding was that shifting from a JSON-based agent structure to a code-based execution model significantly improved performance. When Hugging Face built the first version of Open DeepResearch, it structured AI-driven actions using predefined JSON steps, which limited flexibility.
After switching to a code-based approach, its GAIA score improved from 33% to 55.15%, demonstrating the advantage of more dynamic execution models.
Technical Limitations and Areas for Improvement
Despite promising results, Open DeepResearch still faces several limitations. One of the biggest challenges is the context window constraint, which limits the amount of information the model can process in a single research task. Users testing the tool have reported that it often runs into token overflow errors, causing incomplete responses or failures in handling lengthy investigations.
In our own test we could not get Open DeepResearch running for now in the test environment, but we will try again in the coming days as the project advances.
Token efficiency also remains a critical area for improvement. Several testers reported instances where Open DeepResearch failed to complete tasks due to reaching memory limits. Hugging Face is actively working on solutions, including more efficient text compression and external memory storage options to handle larger queries.
OpenAI’s Deep Research and the Closed AI Debate
OpenAI’s decision to make Deep Research a premium feature for now has intensified debates over access to AI-powered research tools. The move is part of a broader shift in AI development, where companies are restricting access to their most powerful models.
While OpenAI has justified this approach by citing resource costs and security concerns, critics argue that keeping AI-driven research tools behind costly paywalls creates barriers for independent researchers, journalists, and academics.
Hugging Face aims to provide a transparent, open-source research assistant and aims to counterbalance the growing privatization of AI development.
Similar projects, such as the development of DeepSeek, have sought to provide open alternatives to proprietary AI models. However, as seen in the case of DeepSeek R1, not all open AI initiatives come without controversy.
As Hugging Face continues to refine Open DeepResearch, key priorities include expanding browsing capabilities, integrating multimodal AI to analyze images and video, and optimizing memory usage to prevent token limit issues. The company is also exploring alternatives to OpenAI’s o1 model to reduce reliance on proprietary infrastructure.
Independent developers are also building their own AI research agents, with projects from Jina-AI, mshumer with a project called OpenDeepResearcher, and Aomni Co-founder & CEO David Zhang (dzhng) with another open source implementation of OpenAI’s new Deep Research agent.
