KV-Cache Compression and Agent Serving Economics

DeepSeek-V4.1-Flash reports one-million-token context support through asymmetric prefill/decode activation, cross-layer reuse, FP4 caching, and bounded replay, while Token Pooling reports 45–50% lower multi-vector retrieval footprint. DeepSeek elastic-compute discussion adds large CPU and memory-pool capacity signals, but independent quality, latency, interoperability, and TCO validation is still missing.