DeepSeek Open-Sources TileKernels Operator Library, Supporting NVIDIA Blackwell Architecture

The GitHub page reveals that DeepSeek has open-sourced the high-performance GPU operator library TileKernels, developed based on TileLang. This library has been deeply optimized for the training and inference of large language models (LLMs), with operator performance approaching the limits of hardware computational intensity and memory bandwidth. TileKernels includes MoE routing, FP8/FP4 quantization, and various fused operators, and has already been deployed in DeepSeek's internal environment. The library currently supports NVIDIA SM90 and the latest SM100 (Blackwell) architecture, with runtime requirements of CUDA 13.1 or higher.