huggingface/open-r1

Open R1 is an open-source project by Hugging Face, one of the leading AI research platforms, that aims to fully reproduce DeepSeek-R1, a breakthrough reasoning AI model released by a Chinese AI lab in early 2025. DeepSeek-R1 made waves because it demonstrated exceptional reasoning capabilities (particularly in math, coding, and science problems) at a fraction of the cost of competitors like OpenAI's models. However, DeepSeek didn't release all the training details needed to reproduce it. Open R1 is the community's effort to fill those gaps. The project provides the training code, datasets, and step-by-step recipes needed to train your own version of this type of "reasoning model", an AI that shows its thinking process step by step before giving an answer, similar to how a student might show their work on a math problem. The key insight behind these models is that training them to think through problems systematically dramatically improves accuracy on difficult tasks. This is a highly technical research project aimed at AI researchers, machine learning engineers, and teams who want to train their own advanced AI models from scratch. It requires significant GPU infrastructure, the recommended setup is 8 high-end H100 GPUs, and deep familiarity with machine learning training pipelines. For context, Hugging Face has already published several companion datasets generated from this work, and a 7-billion parameter model trained using these techniques that can outperform much larger commercial models on competitive programming benchmarks. The project is actively ongoing and collaborative.