Hugging Face researchers are moving quickly to recreate DeepSeek's R1, the reasoning AI model that drew wide attention after its release. Their project, called Open-R1, is built around a direct idea: if a powerful model is described as open, researchers should be able to inspect how it was made.
The effort is being led by Hugging Face head of research Leandro von Werra and several company engineers. They want to duplicate R1 while making the full process public, including the data used to train it.
Why R1 created so much attention
DeepSeek released R1 last week, and the model quickly became a major topic in AI. On a number of benchmarks, R1 matches and even surpasses the performance of OpenAI's o1 reasoning model.
R1 also reached a wider audience through DeepSeek's chatbot app, which gives users free access to the model. The app rose to the top of the Apple App Store charts, bringing the model beyond research circles and into mainstream conversation.
The reaction was not limited to app rankings. DeepSeek's speed and efficiency in developing R1 led many Wall Street analysts and technologists to question whether the U.S. can maintain its lead in the AI race. Hugging Face, however, is focusing on a different question: how open an open model really is.
The gap between open and open source
R1 is open in one important sense: it is permissively licensed, so it can be deployed largely without restrictions. But Hugging Face researchers argue that this does not make it open source by the widely accepted definition.
The reason is that important parts of the model-building process remain unavailable. DeepSeek did not release an open dataset, experiment details, intermediate models, training code, or training instructions. That makes the model harder to replicate, study, and modify.
Elie Bakouch, one of the Hugging Face engineers working on Open-R1, told TechCrunch that the missing pieces make further research difficult. He described the goal as more than transparency, saying that fully opening the architecture is about unlocking the model's potential.
That distinction matters for AI reasoning models because researchers do not just want to run a model. They want to understand how it behaves, why it behaves that way, and how its behavior can be steered in sensitive areas.
What makes reasoning models different
R1 is described as a reasoning model, which means it effectively fact-checks itself as it works through a problem. That process can help it avoid some of the pitfalls that often affect typical models.
There is a tradeoff. Reasoning models usually take seconds to minutes longer to reach solutions than non-reasoning models. The benefit is that they tend to be more reliable in domains such as physics, science, and math.
For researchers, that makes the training process especially important. If a model is more reliable because of the way it was trained, then understanding the dataset and process is central to improving it. Bakouch said having control over the dataset and process is critical for deploying a model responsibly in sensitive areas and for understanding and addressing biases.
How Open-R1 plans to replicate DeepSeek's work
The Open-R1 project aims to replicate R1 in a few weeks. Hugging Face plans to use its Science Cluster, a dedicated research server with 768 Nvidia H100 GPUs, as part of the effort.
The team intends to use the Science Cluster to generate datasets similar to those DeepSeek used to create R1. It is also asking for help from the AI and broader tech communities through Hugging Face and GitHub, where Open-R1 is being hosted.
Von Werra told TechCrunch that the team needs to make sure the algorithms and recipes are implemented correctly. He said that kind of challenge fits a community effort because it brings many people to the same technical problem.
Early interest has been strong. The Open-R1 project collected 10,000 stars in just three days on GitHub. Stars are a way for GitHub users to show that they like a project or find it useful.
What success could mean for open AI
If Open-R1 succeeds, Hugging Face researchers believe it could give AI researchers a foundation for building the next generation of open source reasoning models. The project is not only trying to create a strong replication of R1. It is also trying to create a usable training pipeline that others can study and extend.
Bakouch argued that open source development is not a zero-sum game. In his view, shared methods can benefit frontier labs, model providers, and independent researchers because they can all build on the same innovations.
The project also enters a debate about the risks of open source AI. Some AI experts have raised concerns about potential abuse, but Bakouch said he believes the benefits outweigh the risks.
He also said that once the R1 recipe has been replicated, anyone who can rent some GPUs can build a variant of R1 with their own data. That would further diffuse the technology and challenge the idea that only a handful of labs can make progress at the frontier of AI.
For Hugging Face, Open-R1 is a test of what openness should mean in a fast-moving AI market. A permissive license allows broad use. A fully visible training process gives researchers something deeper: a way to verify, reproduce, and improve the work.