IBL News | New York
Hugging Face researchers, led by Leandro von Werra, launched Open-R1, a model that seeks to build and fully open-source a duplicate of DeepSeek’s R1, as this Chinese version uses some tools that are not publically released.
Their goal is to replicate R1 in a few weeks, relying in part on Hugging Face’s Science Cluster, a dedicated research server with 768 Nvidia H100 GPUs.
This team of Hugging Face engineers plans to solicit help from AI communities on Hugging Face and GitHub, where the Open-R1 project is being hosted, to build a training pipeline.
Open-R1 comes a week after not-so-open DeepSeek made public its R1 reasoning AI model. It brings all its components, including the data used to train the model, as open-source.
“The R1 model is impressive, but there’s no open dataset, experiment details, or intermediate models available, which makes replication and further research difficult,” Elie Bakouch, one of the Hugging Face engineers on the Open-R1 project, told TechCrunch.“Having control over the dataset and process is critical for deploying a model responsibly in sensitive areas, and it helps with understanding and addressing biases in the model.”
“We need to make sure that we implement the algorithms and recipes correctly, and this is something a community effort is perfect at tackling, where you get as many eyes on the problem as possible,” said Von Werra.
The Open-R1 project attracted 10,000 stars in just three days on GitHub, noting much interest in the project.
“We’re really excited about the recent open source releases that are strengthening the role of openness in AI. It’s an important shift for the field that changes the narrative that only a handful of labs can progress and that open source is lagging.”
Funded by a Chinese quantitative hedge fund, DeepSeek’s R1 matches — and even surpasses — the performance of OpenAI’s o1 reasoning model in several benchmarks.
Being a reasoning model helps it avoid pitfalls that generally trip up models by effectively fact-checking facts, although it takes seconds to respond. The upside is that they tend to be more reliable in physics, science, and math domains.
