Moonshot AI Open-Sources FlashKDA: CUTLASS Kernels for Kimi Delta Attention with Variable-Length Batching and H20 Benchmarks

Moonshot AI open-sourced FlashKDA under an MIT license as a CUTLASS-based CUDA kernel for Kimi Delta Attention.

The company says it delivers 1.72× to 2.22× faster prefill than flash-linear-attention on NVIDIA H20 GPUs.

FlashKDA also adds variable-length batching support, making deployment easier for real-world workloads.

The release strengthens Moonshot AI’s Kimi Linear stack for more efficient long-context LLM inference.