Cursor open-sources its MoE training megakernel MoK
Mixture-of-Kittens fuses all MoE communication and compute into one deterministic Blackwell kernel, lifting Cursor's training throughput 1.41x.
One kernel for the whole MoE layer
Cursor released Mixture-of-Kittens on August 4, an Apache-2.0 training megakernel that collapses every communication and computation step of a Mixture-of-Experts layer into a single fused kernel. Rather than alternating between expert compute and all-to-all token routing — the pattern that makes MoE training bandwidth-bound — MoK overlaps them inside one launch.
The code targets Nvidia Blackwell SM100/SM103 parts, meaning GB200 NVL72 or GB300 NVL72 racks, and requires Python 3.12+, PyTorch 2.10+ and CUDA 13.0+. On a single-layer MXFP8 forward-pass benchmark Cursor reports up to 2.37x the throughput of the strongest public baseline. The number that matters more is end to end: across 512 GPUs, replacing its previous DeepEP-based path lifted training throughput 1.41x, from 760.9 to 1,070.2 tokens per second per GPU. Cursor says MoK is already carrying training across tens of thousands of its GPUs.
Determinism as a feature
MoK is bit-wise deterministic. That is unusual for fused collective kernels and it is deliberate: on-policy reinforcement learning post-training breaks in subtle ways when the training and sampling paths disagree numerically, and non-determinism makes regressions almost impossible to bisect. A deterministic MoE kernel makes RL runs reproducible and debuggable, which is arguably worth more to a lab than the raw speedup.
Why it matters
A 1.41x throughput gain is a direct 30 percent cut in the cost and wall-clock of every MoE pretraining and post-training run on Blackwell racks — the substrate nearly every frontier model now trains on. Publishing it under Apache 2.0 hands that saving to anyone with NVL72 access, narrowing one of the gaps between well-resourced labs and everyone else, since this class of kernel work usually stays internal. It also says something about Cursor. A company known for a code editor is now shipping frontier-grade training infrastructure, evidence that its shift into training its own models is real rather than a wrapper strategy — and a reminder that hardware-specific kernel engineering, not just architecture, is where a growing share of training efficiency now comes from.