DeepSeek R1 is now a peer-reviewed AI model

Chinese AI firm DeepSeek has published a peer-reviewed paper in the scientific journal Nature, describing its large language model R1, which is designed to handle reasoning tasks such as maths and coding. The company said the system was not trained on the output of rival models, addressing speculation about its data sources.

R1 is an “open weight” model, meaning its code is available for download, and it has become the most downloaded model on the Hugging Face platform, with more than 10 million downloads, states an article by Nature on the paper.

According to the paper, training R1 cost about $294,000, in addition to roughly $6 million spent on the base model, significantly less than the estimated tens of millions needed for comparable systems.

The model was trained mainly on Nvidia’s H800 chips, which are restricted from export to China under US controls. Researchers said R1 was built using reinforcement learning, rewarding correct answers rather than imitating human reasoning patterns. The paper also notes adjustments made during peer review, such as clarifying training data and safety considerations.

This marks the first time a major language model has been formally published through peer review, setting a precedent for transparency in AI development. You can read the Nature article here.