DeepSeek open-sources DeepInfer inference system

DeepSeek is framing DeepInfer as its production stack rather than a side project: the system is used in its own AI services, and the open-source release collects its core functions. The company says it is sharing the stack to help the community lower compute costs and reduce energy use.

The framework did not appear from scratch. It builds on vLLM, a widely used system for serving AI models at scale, and borrows ideas from SGLang and TensorRT LLM — the same set of tools DeepSeek names as the common options for large-scale inference. The release supports running DeepSeek-V4.1-Flash and includes optimized inference features and an acceleration technique called DSpark.

The companion release, deepseek-recipe, targets a different layer: it handles token encoding and decoding aligned with the official DeepSeek API. For third-party API providers, that may make it easier to offer behavior matching DeepSeek's own API, which is likely the practical value beyond the inference code itself.