CacheBlend is released as part of LMCache and vLLM Production Stack. It enables reuse of KV caches from non-prefix text chunks for RAG workloads, aiming for near-100% KV cache hit rate while selectively recomputing cross-attention to preserve output quality.