DeepSeek open-sources DeepInfer, the inference engine behind its API, under Apache 2.0
DeepSeek has open-sourced DeepInfer, its internal inference framework, under the Apache License 2.0. Inference is the serving layer that lets DeepSeek's V4.1-Flash model generate answers quickly at scale, and the code is now public.
The Decoder reports DeepInfer builds on vLLM with optimized compute kernels, distributed processing across multiple chips and faster text-generation techniques. DeepSeek says code for its future models will also ship as open source.
Alongside the engine, DeepSeek released deepseek-recipe, which replicates the token processing of the official DeepSeek API so self-hosted setups behave the same way as the hosted service.
Coverage in The New Stack notes the release gives Huawei equal billing with Nvidia: DeepInfer supports Huawei Ascend hardware alongside Nvidia GPUs, a notable choice given US export controls on advanced chips.
DeepSeek framed the release on X as a way 'to help the community lower compute costs, reduce energy use, and make AI more affordable.' Its models are already priced well below Western rivals, and a free serving stack adds pricing pressure on OpenAI and Anthropic.
What it means for buyers: teams self-hosting DeepSeek models can now match the official API's serving behavior and token processing. Hardware support lists matter before planning a cluster; check the repo for your GPU generation. DeepSeek's code carries the usual considerations for any foreign-origin infrastructure software.