DeepSeek just open-sourced its inference engine

The repository covers one model, the only supported model is DeepSeek-V4.1-Flash, and it targets Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. Building it requires Linux, GCC 13.0 or higher, Python 3.12, and either CUDA 13 for NVIDIA hardware or CANN 9.2 for Ascend hardware. A complete production service still needs an external KV cache system, PD disaggregation, and load balancing across replicas. Reported production figures include throughput of 155 trillion tokens per day per cluster and an input cache hit rate of 96%. The framework was built mostly on top of vLLM, with ideas borrowed from SGLang and TensorRT-LLM. A companion token codec, deepseek-recipe, ships alongside it. Discussion centers less on the technical merits and more on why DeepSeek keeps open-sourcing its secret sauce. For an AI company, the practical value is a permissively licensed reference for MoE serving and speculative decoding, plus a second hardware path beyond Nvidia. DeepSeek says yes but the release is a subset. Teams should benchmark before assuming parity with the hosted API.