DeepSeek Open-Sources DeepInfer Framework to Lower AI Compute Costs

Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. has open-sourced DeepInfer, a high-performance inference framework used for its own production services. Released under the Apache License 2.0, the software is designed to reduce energy consumption and computing costs for large-scale AI models, specifically supporting the DeepSeek-V4.1-Flash model.

The framework is compatible with Nvidia Corporation SM100 Blackwell GPUs and Huawei Technologies Ascend 950 NPUs. DeepInfer incorporates asynchronous scheduling, deterministic sampling, and speculative decoding via DSpark to increase throughput. While the company released the core framework and the deepseek-recipe library, it withheld load-balancing code and key-value cache management due to their integration with internal infrastructure.

This release comes as the United States Department of Homeland Security ordered federal civilian agencies to remove DeepSeek products and related third-party services from their systems. Market analysts suggest that by lowering inference costs, the tool may challenge investor assumptions regarding the necessity for continuous, high-paced spending on data-center infrastructure and GPUs.