tl;dr: Kimi Linear is a hybrid attention architecture combining Kimi Delta Attention (KDA)—a gated linear attention variant with hardware-efficient Diagonal-Plus-Low-Rank transitions—with Multi-Head Latent Attention (MLA) in a layerwise mix. A 48B-parameter (3B active) model reportedly beats full MLA attention across short-context, long-context, and RL tasks while cutting KV cache by up to 75% and delivering 6x faster decoding at 1M context. The team has open-sourced the KDA kernel, vLLM integration, and model checkpoints.