DeepSeek has published DeepInfer, its in-house inference tool, as open source on GitHub, announced through the company's X account. It currently supports only V4.1 Flash, the faster and cheaper variant of its flagship, with DeepSeek saying further models will be added as they become public.
Releasing the serving layer is a different act from releasing weights. Open weights tell you what the model is; an inference engine is how you run it economically - batching, scheduling, memory management, the unglamorous work that decides whether a model costs ten cents or a dollar to serve. Labs generally keep that. DeepSeek's stated reason is to help the community lower compute costs and energy use.
A pricing note worth care. Coverage of the release cites OpenRouter listing V4.1 Flash at $0.035 per million input and $0.12 per million output, against GPT-6 Astra at $10 and $50. DeepSeek's own published rate for V4.1-Flash-1017 is $0.30 input and $1.20 output at peak, roughly half that off-peak. Those are an order of magnitude apart, so the OpenRouter figure is either a promotional or third-party rate rather than DeepSeek's list price. Either way the gap against frontier pricing is large - just check which number applies to you before building a cost model on it.