Leandro von Werra, Head of Research at Hugging Face, along with several other engineers at the company, launched the Open-R1 project, which aims to create a duplicate of the R1. The goal of this project is, among other things, to disclose the data used for training.
DeepSeek's R1 model is technically already "open source" and anyone can use it without any restrictions. However, this model does not fall under the generally accepted definition of open source software.
Researchers from Hugging Face told TechCrunch that hidden elements make it difficult to replicate and further research the model. In particular, the Chinese startup does not disclose the training dataset, details of experiments, and intermediate models.
The American company plans to create a replica of R1 within a few weeks. For this purpose, a special research server Science Cluster equipped with 768 NVIDIA H100 GPUs will be used.
Another goal of Huggin Face with Open-R1 is that, if the project is successful, users and developers will be able to create the next generations of reasoning LLMs, including open source.