DeepSeek Releases DeepGEMM: Open-Source GPU Matrix Acceleration Library for AI Models

DeepSeek has open-sourced DeepGEMM, a library that accelerates GPU matrix operations used by large language models, with support for NVIDIA Hopper and Blackwell.
DeepSeek has released DeepGEMM, an open-source library that accelerates GPU computations underlying AI models. It optimizes matrix operations used by large language models and improves efficiency on NVIDIA Hopper and Blackwell GPUs.
The library supports FP8, FP4, and BF16. It includes optimizations for Mixture-of-Experts architectures, overlaps inter-GPU data exchange with computation, and compiles kernels at runtime, eliminating the need for a separate CUDA build during installation.
DeepSeek states that DeepGEMM reaches performance at or above specialized expert-tuned libraries for a number of matrix sizes.
Why it matters
AnalysisEfficient matrix kernels can reduce training and inference costs for LLMs, especially on modern NVIDIA GPUs and Mixture-of-Experts architectures.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
DeepSeek open-sources part of its chat interface
DeepSeek has released the source code for a portion of its chat UI, including web search, deep thinking, and conversation history features.
Open Weights vs Truly Open Source AI
A legal and engineering analysis of the fundamental difference between accessible numerical weights of models (Llama, DeepSeek) and fully open projects with source data, code, and architecture (OSI Standard).
Codync: open-source Rust tool turns AI agents into chatbots
Codync is a free Rust project that wraps more than 40 local AI agents into chatbots, enabling control from iPhone, Mac, Windows, Linux, or the terminal.
Frontier Models
The most powerful class of artificial intelligence at the forefront of global research (Claude 3.7 Sonnet, OpenAI o3/GPT-4.5, Gemini 2.0 Pro), defining the limits of modern reasoning, autonomy, and coding capabilities.