DeepSeek opens source of DeepInfer framework to promote model accessibility

On October 20, 2026, DeepSeek announced the full open-sourcing of its internal inference framework, DeepInfer, as part of its ongoing efforts to make large models more accessible. The company emphasized that while model capabilities are defined by the model itself, the efficiency of inference frameworks significantly impacts the cost and energy consumption of large-scale AI services. By sharing its deployment experience, DeepSeek aims to help developers and organizations optimize large-scale inference performance, reduce computational costs, and enhance energy efficiency. The open-sourced DeepInfer supports distributed inference, speculative decoding, and compatibility with NVIDIA and Huawei Ascend platforms. DeepSeek also released the deepseek-recipe component, aligning with its official API for consistent community experience. DeepInfer is available at https://github.com/deepseek-ai/DeepInfer, and deepseek-recipe at https://github.com/deepseek-ai/deepseek-recipe.