DeepSeek open-sources DeepInfer inference framework and tokenizer library

The engine powers DeepSeek-V4.1-Flash in production with distributed inference, custom kernels, speculative decoding, and structured outputs.

DeepInfer includes SWA Bounded Replay alongside its optimized engine and kernels
Inference code and optimizations for future DeepSeek models will be added directly to the repository
DeepSeek also open-sourced deepseek-recipe to replicate the official API's token encoding and decoding logic in self-hosted deployments

Engineers self-hosting DeepSeek models can use the exact serving stack and token-processing logic DeepSeek runs in production.

Sources