DeepSeek open-sourced three infrastructure repos in four days.

DeepSelect is the TopK kernel behind the indexer in DeepSeek Sparse Attention, and DeepSeek says it runs 2-20x faster than torch.topk. DeepJIT is a header-only C++20 runtime that compiles kernels at runtime, with one interface for both NVIDIA CUDA and Huawei Ascend. deepseek-recipe converts Messages, Chat Completions and Responses requests into V4 and V4.1 prompts for self-hosting. The Ascend support tells you the most: DeepSeek is building its kernels to run on Huawei hardware as well as NVIDIA.