Hugging Face researchers are trying to build a more open version of DeepSeek’s AI ‘reasoning’ model

Jan 29, 2025

Hugging Face aims to open-source DeepSeek's AI model to enhance transparency and collaboration.

Hugging Face researchers are trying to build a more open version of DeepSeek’s AI ‘reasoning’ model

: Hugging Face researchers, led by Leandro von Werra, launched Open-R1 to replicate DeepSeek's R1 AI model and make it fully open source. They aim to build a duplicate of R1 due to its 'black box' release approach, lacking open data and models. By collaborating with the AI community and leveraging 768 Nvidia H100 GPUs, Hugging Face seeks to create similar datasets and training pipelines. The project received 10,000 GitHub stars in three days, reflecting strong community interest in open-source AI models.

Hugging Face, led by Leandro von Werra, has initiated Open-R1, a project focused on replicating DeepSeek's R1 AI model to foster 'open knowledge.' By open-sourcing all components, including training data, the researchers target DeepSeek's limited transparency approach, which hampers further study and development.

DeepSeek's R1, a popular reasoning model, captured attention after topping the Apple App Store charts and outperforming some benchmarks of OpenAI's o1 model. However, it remains partially closed with undisclosed components, prompting concerns from researchers about its opaque development process and challenges in addressing potential biases.

Utilizing Hugging Face’s Science Cluster comprising 768 Nvidia H100 GPUs, Open-R1 plans to replicate R1’s training pipeline via community collaboration on platforms like GitHub. With rapid community engagement, including 10,000 GitHub stars in three days, the project aspires to enable broader AI innovation, emphasizing open source as a means to benefit the entire field.