DeepSeek performs excellently on multiple code benchmarks (such as HumanEval, MBPP, and LiveCodeBench), with some metrics approaching or even surpassing GPT-4-level models.