DeepSeek-Math 7B is an open-source mathematics large model with 7B parameters, whose RL version with tool invocation can achieve approximately 60% accuracy on the MATH benchmark, approaching GPT-4 level.