Cursor Open-Sources Mixture-of-Kittens, the Fully Deterministic MoE Training Megakernel Behind Composer
Cursor Research open-sourced Mixture-of-Kittens (MoK) on 2026-08-04 — the production MoE training megakernel that powers Composer training across tens of thousands of GPUs. MoK fuses all MoE communication and computation into a single, fully deterministic kernel built from first principles for NVL72 racks, and Cursor claims state-of-the-art performance against publicly available implementations. The repo (cursor/mixture-of-kittens, Apache-2.0, Python) was created July 29 and is at 326 stars with commits as recent as this morning. Determinism is the underrated part: reproducible MoE training runs make training-bug bisection tractable, which is why Cursor explicitly pitches it as agent-modifiable for other hardware platforms.
↳ Follow the thread