DeepSeek has released its DeepInfer inference technology as open source to improve performance and reduce costs. The company stated that the stack powers its DeepSeek-V4.1-Flash model, offering high throughput and low latency. The release includes high-performance kernels, an optimized inference engine, distributed inference, and speculative decoding capabilities.
The company also open-sourced deepseek-recipe, a component for token encoding and decoding that aligns with the processing logic of the official DeepSeek API. This move aims to help developers and cloud providers reduce compute expenses and energy consumption. DeepSeek intends to continue collaborating with the community to enhance large-scale inference efficiency.
The announcement follows a recent order from the U.S. Department of Homeland Security, which prohibited federal civilian agencies from acquiring DeepSeek products. The agency cited security concerns related to the company's status as a privately owned entity based in China, noting that its models present multiple threat vectors.
