Hugging Face is building Open R1, a fully open source version of DeepSeek R1 AI model

Hugging Face has announced it's building Open R1, an open reproduction of the recently launched DeepSeek R1 model. While DeepSeek R1's model weights are accessible, the datasets and training code remain undisclosed. Open R1 aims to fill these gaps, enabling the research and industry community to develop similar or superior models.

The initiative focuses on reconstructing DeepSeek R1's data and training processes, validating its claims, and advancing open reasoning models. Hugging Face seeks to enhance transparency in reinforcement learning's role in reasoning and provide reproducible insights to the open-source community.

Open R1 is not just about replicating results, but also about sharing valuable insights. By documenting effective and ineffective strategies, Hugging Face hopes to prevent others from expending unnecessary time and compute resources on unproductive approaches.