Cursor open-sources Mixture-of-Kittens, a deterministic MoE training megakernel for NVL72 racks

On August 4, 2026, Cursor released Mixture-of-Kittens (Apache 2.0), a fully deterministic mixture-of-experts training megakernel for GB200/GB300 NVL72 racks that fuses all MoE compute and communication into a single kernel with no CPU–GPU synchronization. Cursor says it delivers up to 2.37× faster MXFP8 forward passes and lifted its production Composer training throughput 1.41×.

What it means An application company open-sourcing frontier-grade training kernels signals how far serious model infrastructure has spread beyond the big labs.