DeepSeek-R1-Zero required 30 times less computing power

Researchers have revised the cost estimate for training the DeepSeek-R1-Zero model. According to updated data, the reinforcement learning (RL) process required approximately 2e22 FLOPs, which is 30 times less than initial projections.

The adjustment of these figures was made possible by recalculating the batch size and refining the Model Flops Utilization (MFU). Initial estimates did not account for the specifics of load distribution during RL training, which led to a significant overestimation of the required resources in early reports.

This clarification demonstrates that modern reinforcement learning methods can be significantly more cost-effective than previously assumed. For developers, this means the ability to plan infrastructure more accurately when scaling similar models, and it also confirms the high efficiency of current optimization algorithms used at DeepSeek.