Moonshot AI Open-Sources FlashKDA: 1.7–2.2x Attention Speedup for Kimi Linear's Delta Attention Architecture

Moonshot AI / GitHub·medium signal

Moonshot AI released FlashKDA, a CUTLASS-based kernel implementation of Kimi Delta Attention delivering 1.72x–2.22x prefill speedup over flash-linear-attention on NVIDIA H20 GPUs. Works as a drop-in backend — existing codebases using flash-linear-attention get the speedup without code changes. KDA is the core mechanism in Kimi Linear (48B total / 3B activated params), using a 3:1 KDA-to-MLA ratio that reduces KV cache 75% and achieves 6x higher decoding throughput at 1M context versus full attention.