ThunderKittens
libraryHazy Research's tile-primitive embedded DSL for writing fast AI kernels (3.5K★), matching or beating SOTA attention and linear-attention kernels with far less code; TK 2.0 (2026) and HipKittens (AMD CDNA port, 2025) extend the family, and the Megakernels line ("no bubbles" single-kernel inference) applies the same philosophy to serving. The successor systems agenda to FlashAttention.