Illustration of a kitten-shaped kernel, representing Cursor's open-source Mixture-of-Kittens megakernel for MoE training.

Editorial illustration for Cursor Open-Sources Mixture-of-Kittens, a Megakernel for MoE Training

Cursor Open-Sources Mixture-of-Kittens, a Megakernel for...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

August 4, 2026 • 3 min read

Cursor Research put its mixture-of-experts training code out in the open this week, releasing Mixture-of-Kittens, or MoK, on GitHub under an Apache-2.0 license. The kernel sits behind Composer, Cursor's own model line, and has already run training jobs across tens of thousands of GPUs. The pitch is throughput: Cursor's team reports up to 2.37x the speed of the strongest public baseline, achieved by fusing every MoE communication and computation step into one deterministic kernel instead of stitching separate operations together.

None of that comes cheap on hardware. MoK needs NVIDIA Blackwell SM100 or SM103 chips, meaning GB200 NVL72 or GB300 NVL72 racks, plus Python 3.12+, PyTorch 2.10+, and CUDA toolkit 13.0+. Its inter-GPU buffers depend on PyTorch symmetric memory, a detail that rules out anyone without rack-scale NVLink access.

That narrows the real user base to frontier labs, well-funded model startups, GPU neoclouds, and national computing centers. Eight-GPU shops need not apply.

The bottleneck MoK targets didn't show up where Cursor first expected it.

Cursor Research has open-sourced Mixture-of-Kittens (MoK), the mixture-of-experts training megakernel behind its Composer models. MoK fuses every MoE communication and computation step into a single deterministic kernel. Cursor team reports up to 2.37x higher throughput than the strongest public baseline.

Cursor Open-Sources Mixture-of-Kittens (MoK): A Deterministic MoE Training Megakernel for GB300 NVL72 Racks - MarkTechPost

Why this matters MoK is a useful data point on where MoE training bottlenecks actually live once you clear the compute problem: the communication layer. Cursor's own numbers, a 2.37x throughput gain over the strongest public baseline, are self-reported and tied to Composer, so we'd want independent benchmarks before treating that figure as gospel. Still, the fact that a company built its own MXFP8 and NVFP4 kernels, then found the MoE layer eating over half of training time anyway, tells us something about the state of the tooling everyone else is using.

The Apache-2.0 license makes MoK genuinely inspectable, which is more than most lab infrastructure releases offer. The catch is the hardware floor: NVIDIA Blackwell SM100, GB300 NVL72 racks, tens of thousands of GPUs. That's not a kernel you drop into a university cluster or a startup's rented A100 pool.

For researchers, it's worth reading as a blueprint for where MoE training is headed. For most teams, it's closer to a preview of problems they'll hit once they can afford the hardware to have them.

Ship an AI product this weekend — no engineers required.

Structured, in-depth lessons on the exact no-code tools — not scattered tutorials.

The exact platforms, taught in depth
Build real, working projects
Our honest review + a reader discount

Read the review →

Cursor Open-Sources Mixture-of-Kittens, a Megakernel for...

Further Reading

Ship an AI product this weekend — no engineers required.

Latest News

Pulitzer Winners Set Record for AI Use in Investigations