DeepSeek has made DeepInfer, its in-house inference tool, freely available on GitHub, putting a piece of its own production stack in the open as it pushes to make AI cheaper to run. The framework is built today to work only with DeepSeek’s V4-1 Flash model, a faster, more efficient and cheaper version of its flagship system.
The timing matters because inference is where a trained model starts doing the work people actually pay for: turning prompts into outputs. It is also one of the most expensive parts of the AI lifecycle, and a report from McKinsey & Company earlier this year said that by the end of the decade most data-center compute would be devoted to inference rather than training new models. DeepSeek is aiming squarely at that cost center, saying it is sharing its production stack to help the community lower compute costs, reduce energy use and make AI more affordable.
The pricing gap is stark. DeepSeek’s V4.1-Flash is listed by OpenRouter at $0.035 per million input tokens and $0.12 per million output tokens. OpenAI’s GPT-6 Astra is listed at $10 per million input tokens and $50 per million output tokens. That difference explains why any tool that helps trim inference overhead gets attention so quickly: the room to save starts before a customer ever changes model providers.
DeepInfer could push those bills down further because it uses multiple inference techniques, including speculative decoding. In plain terms, that means the system tries to speed generation by making and checking likely next-token guesses instead of waiting for every token to be produced one by one. DeepSeek did not quantify how much money that might save in practice, and it did not give a timetable for when other models will be added to the framework, only saying future models will arrive as they become publicly available.
There is also a hard edge to the release. On Monday, the Trump administration banned federal civilian agencies such as the State Department from using DeepSeek’s AI tools, even as the company opened DeepInfer to wider use. The rule does not cover private companies or other organizations, which leaves the commercial side of the market open even as the federal government moves in the opposite direction.
DeepSeek, a Hangzhou-based startup that badly rattled Silicon Valley earlier last year with the release of its R1 model, is now making the case that cheaper inference is where the next battle will be fought. The open question is not whether demand for lower-cost AI is real. It is how much DeepInfer can actually cut the bill once it leaves DeepSeek’s own stack and lands in other hands.
