MoK is a material training-efficiency contribution for specialised GB300 deployments, but the reported performance advantage remains insufficiently independently validated in this source.
Limits of the evidence: The item is independent reporting of Cursor's claims, provides no benchmark methodology or independent reproduction, and states that the kernel requires scarce GB300 NVL72 capacity.
What to watch
Taken from the first report; the whole story has not been re-assessed yet.
Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks
MarkTechPostAug 4, 2026First report
No revisions yet.
No maturity ladder applies to this desk.
Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original sources are authoritative.