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

Cursor Research just released Mixture-of-Kittens, or MoK, the megakernel that trains its Composer models across tens of thousands of GPUs. It's a single, deterministic piece of software that fuses every communication and compute step of mixture-of-experts training into one unit, rather than juggling them separately. Cursor says it beats the best public baseline by as much as 2.37x on throughput, and it's not a lab demo — this is the actual code running Composer's production training runs right now, dropped on GitHub under Apache-2.0.

The backstory here is that Cursor already solved the compute half of this puzzle with custom MXFP8 and NVFP4 kernels for MoE inference. What they hadn't fixed was communication, and it turned out communication was eating more than half of total training time once you scale onto GB300 NVL72 racks — 72 GPUs wired into a single NVLink domain. That density opens the door to tight overlap between GPUs, but it also exposes a weakness: the Grace CPUs bolted onto those racks are comparatively sluggish, so every CPU-GPU handshake becomes a tax you want to avoid.

Three engineering choices do most of the work. First, MoK flips the usual playbook by pulling data instead of pushing it for forward dispatch, because Cursor's benchmarks showed push-based transfers like DeepEP's leave one NVLink direction nearly idle under imbalanced expert loads. Pull dispatch alone cut signaling latency from 103 microseconds to 18. Second, the team picked an overlap granularity that sits between DeepEP's coarse chunks and Comet's fine-grained approach, tuning it so each expert-grouped GEMM gets at least two full waves of streaming multiprocessors — for Kimi 2.5-shaped workloads that works out to a floor of 2,368 tokens. Third, instead of letting the CPU manage buffer sizing or dropping tokens when things get crowded, MoK cycles a small ring buffer of a few hundred megabytes entirely on the GPU side, walked in reverse during backward passes to cut down on replaying activations.

The numbers back up the pitch. Tested across shapes like Kimi K2.7 Code, GLM-5.2, and DeepSeek-V4-Pro on a single NVL72 rack, MoK hit 2.37x on MXFP8 forward passes and still delivered solid gains — 1.78x, 1.92x, 1.58x — on the other precision and direction combos. Scaled out to 512 GPUs across multiple racks, per-GPU throughput climbed from 760.9 to 1,070.2 tokens per second, a 1.41x jump that's harder to dismiss than a cherry-picked microbenchmark.

The catch is the entry price. MoK demands Blackwell SM100 or SM103 silicon, meaning GB200 or GB300 NVL72 racks specifically, plus Python 3.12, PyTorch 2.10, and CUDA 13.0 at minimum. That's not a stack you spin up on a handful of workstation GPUs. It's built for frontier labs, well-funded startups, GPU neoclouds, and national compute centers — anyone else can admire the code on GitHub and wait for the hardware to trickle down.