DeepSeek has released TileKernels, a GPU kernel library designed for large model training and inference, under the MIT license. Announced on April 23, TileKernels is written in TileLang, a Python-based domain-specific language developed by tile-ai for high-performance GPU kernels. The library includes six categories of kernels, such as MoE gating, quantization, and Engram gating, with some components already deployed internally. This release marks the first public disclosure of DeepSeek's proprietary Engram and Manifold HyperConnection components. The library requires NVIDIA SM90 or SM100 architecture GPUs, CUDA Toolkit 13.1 or higher, and PyTorch 2.10 or higher.