DeepSeek Opens Its Production Inference System DeepInfer
Open-sourcing the stack the company actually runs shifts the practical question for third parties from whether they can reproduce DeepSeek's API behavior to whether they can do it without the KV cache the release leaves out.
Reporting from 1 source: GIGAZINE.
DeepSeek published DeepInfer, the inference system its own AI services use, with core functionality packaged for outside deployment. It is built on vLLM with ideas from SGLang and TensorRT LLM, runs DeepSeek-V4.1-Flash, and supports the DSpark acceleration technology, but a KV cache system is not implemented. DeepSeek also released deepseek-recipe, which converts requests and outputs between third-party APIs and the official DeepSeek API format.
DeepInfer carries the Apache License Version 2.0 and sits in DeepSeek's GitHub organization. The release notes a gap: no KV cache system is implemented, so what ships is the core functionality rather than the full serving stack.
The companion repository, deepseek-recipe, comes under the MIT License. Its conversion tools let third-party API providers reproduce behavior equivalent to the official DeepSeek API, which matters for anyone reselling or proxying the models.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.