What's this about?
Cursor released MoK, a "megakernel" for training Mixture-of-Experts models on NVIDIA NVL72 clusters. It fuses all the communication and computation of MoE training into a single, fully deterministic GPU kernel, running up to 2.37× faster than the strongest public baselines. Notably, Cursor — known as a code editor — is now publishing frontier *training* infrastructure.
What exactly is a "megakernel"?
Normally a GPU runs a training step as dozens of separate operations (kernels), each launched and coordinated by the CPU, with data shuffled between experts across the network in between. A megakernel welds all of that into one giant GPU program, eliminating launch overhead and idle waiting. "Fully deterministic" means the same inputs always give bit-identical outputs — rare and valuable in training, because it makes runs reproducible and bugs findable.
What's the community saying?
- Surprise at the source: a code-editor company shipping a competitive MoE training kernel signals Cursor is training its own frontier models, not just wrapping others'.
- The determinism angle draws engineers — reproducible training is a known pain point, and few open kernels offer it.
- The 2.37× claim invites the usual "vs. which baseline, on which hardware" scrutiny; NVL72-specific means most can't reproduce it directly.