DeepSeek open-sources DeepInfer; KV-cache management and load-balancing code is not yet fully open

DeepSeek says it is open-sourcing DeepInfer, an inference framework for large-scale cloud services that powers its V4.1 Flash service. It supports distributed inference, speculative decoding, SWA bounded replay and structured outputs. DeepSeek says it is also open-sourcing deepseek-recipe, a component that converts between mainstream LLM API protocols and model token sequences. Why it matters: Read it as a partial systems release: DeepSeek says KV cache management and load-balancing code are not yet fully open-sourced because they are tightly coupled to its internal infrastructure. It is a useful code and architecture reference, not independent evidence that another deployment will match DeepSeek’s service performance.