DeepSeek has open-sourced DeepInfer, the inference framework that powers its DeepSeek-V4.1-Flash model in production. The release is under the Apache 2.0 license, making the code freely available for external use and modification.
DeepInfer is designed for specific hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. This dual support suggests DeepSeek is catering to both mainstream accelerator ecosystems and China's domestic AI chip push.
The framework currently ships with offline inference only, meaning it does not include the components needed for live, online request serving. That limitation is worth noting for developers who want to deploy the model in interactive services rather than batch processing.