DeepGEMM is an open-source matrix multiplication kernel library developed by the Chinese AI company DeepSeek. It is designed to perform GEMM (general matrix multiplication), the dominant computation in training and inference of large language models, efficiently at FP8 precision, and was released in February 2025 during DeepSeek's open-source week.
It supports both standard GEMM for dense models and grouped GEMM that batches per-expert computations in MoE models, and compiles its kernels at runtime. It is one of the libraries on which DeepSeek's inference engine DeepInfer is built.