DeepSeek Open‑Sources DeepInfer Inference Framework and Production Token Codec

Overview

• DeepSeek published DeepInfer and a production token codec called deepseek-recipe on GitHub as public repositories used in its production service.
• The company says it open-sourced the stack to help developers and enterprises lower compute costs and reduce energy use when running large-model inference at scale.
• DeepInfer includes high-performance operators and a tuned inference engine that supports distributed inference, speculative decoding, SWA Bounded Replay, and structured output.
• DeepSeek already runs DeepInfer in its DeepSeek V4.1 Flash service and the framework is compatible with NVIDIA GPUs and Huawei Ascend hardware.
• The repositories include the exact components DeepSeek uses in production and the company has committed to ongoing updates, which could make deploying efficient, private or cloud inference more accessible for organizations.