Cursor open-sources MoE megakernel for GB300 NVL72 racks

Cursor's Mixture-of-Kittens release is the kind of infrastructure drop that says more about where frontier training is heading than another benchmark chart would. The team says the mixture-of-experts layer was consuming more than half of end-to-end training time in some configurations of their Composer model, and prior optimization work had squeezed the compute side without touching the communication side. MoK's answer is to fuse all MoE computation and inter-GPU communication into a single, fully deterministic kernel targeted at NVIDIA's NVL72 racks.

The numbers Cursor reports, posted alongside a blog write-up and coverage from MarkTechPost, are worth taking seriously but as reported, not settled. Against what they call the fastest public baseline, MoK is up to 2.37x faster on the MXFP8 forward pass, 1.78x on the MXFP8 backward, 1.92x on the BF16 forward, and 1.58x on the BF16 backward. In production, on their own Composer training stack across several NVL72 racks, they report a 1.41x jump in end-to-end tokens per second per GPU, from 760.9 to 1,070.2.

What is interesting is less the specific multipliers and more the shape of the work. MoK is hard-locked to NVIDIA Blackwell SM100 and SM103 GPUs, meaning GB200 NVL72 or GB300 NVL72 racks, and it needs Python 3.12, PyTorch 2.10, and CUDA 13.0 or newer. This is not a portable library, it is a bet that the future of large MoE training runs through a specific rack-scale NVIDIA product and that squeezing it requires bespoke kernels rather than framework-level tricks. The authors listed on the release, Stuart Sul, Nash Brown, Henry Wildermuth, William Lin, and Federico Cassano, are effectively publishing the piece of infrastructure that powers Cursor's own Composer training.

The honest caveats are the obvious ones. The speedups are measured against Cursor's chosen baseline on Cursor's model, and the reporting doesn't give you the MoE architecture details, expert counts, or batch settings behind the 1.41x production number, so how it generalizes to other teams' MoE designs is genuinely an open question. What is not in doubt is the direction: at the frontier, the interesting engineering is moving into the kernel, into determinism, and into hardware you have to physically stand up to use.