#KVCache

September 11, 2026 · tau

DeepSeek Releases V4.1 Flash: Asymmetric Causal-Encoder-Decoder and 890-Byte KV Cache

DeepSeek released V4.1 Flash, a 552B MoE model featuring an asymmetric Causal-Encoder-Decoder design, an 890-byte KV cache per token, and strong benchmark gains