Cursor open-sources Mixture-of-Kittens, a GPU training kernel

Cursor, the AI coding startup behind the Composer model, released the full source code for Mixture-of-Kittens, a custom kernel built to speed up training on Nvidia’s newest GB300 NVL72 systems. The company says the tool lifted end-to-end training throughput by 41 percent on its production stack, measured across 512 GPUs spanning several NVL72 racks, compared with the DeepEP-based setup it replaced.

Mixture-of-Kittens targets the mixture-of-experts layer, the architecture used by Composer and by open-weight models including DeepSeek, GLM, Qwen and Kimi. Cursor says that layer had become the dominant training cost, at times consuming more than half of total run time depending on batch size and configuration.

A normal GPU program for this kind of workload launches a separate step for each task: route the tokens, ship them to the chip holding the right expert, run the math, send results back. Each handoff forces the GPU to pause, flush its work and reload data from memory before the next step can start. Cursor fused routing, network transfer and matrix computation into one kernel that keeps data resident on the chip and lets the math for one batch of tokens run while another batch is still crossing the network.

Mixture-of-experts models route each token to a handful of specialized subnetworks. With those experts spread across dozens of GPUs, different tokens must physically travel to different chips before any computation can happen. Cursor’s writeup identifies that shuffle, not the matrix arithmetic, as the real constraint. The core engineering problem was hiding the GPU-to-GPU data movement behind ongoing computation so neither one sits idle waiting on the other.

Isolating communication direction alone produced one of the sharper results. Cursor found that having each GPU pull tokens from their source, rather than the more common approach of pushing them to a destination, cut the wait for a completion signal from roughly 103 microseconds to 18 microseconds and raised NVLink bandwidth use by as much as 29 percent under uneven expert load.

On isolated layer benchmarks, run within a single NVL72 rack at 64-way expert parallelism with 2,048 tokens per GPU, Mixture-of-Kittens beat the fastest of four baselines (NCCL, two DeepEP variants and Nvidia’s own HybridEP) by up to 2.37x on the forward pass in the low-precision MXFP8 format and 1.92x in BF16. On Cursor’s live production stack, running 512 GPUs across multiple GB300 racks, the gain settled at 1.41x, taking tokens processed per second per GPU from 760.9 to 1,070.2.

Every one of those comparisons is Cursor’s own measurement against baselines it selected and ran itself, and the headline production number compares the new kernel only against Cursor’s prior internal pipeline, not against a rival vendor’s system.

Cursor is not a chip maker or a research lab. It sells a code editor and an AI assistant built around Composer. That it now writes deterministic, hand-tuned GPU kernels down to the level of streaming-multiprocessor scheduling says something about where inference and training economics currently sit. Once the expert-routing layer eats half of training time, generic communication libraries stop being good enough, and companies serving their own models at scale face pressure to own the lowest levels of the stack rather than rent them. Any team running frontier-scale training or inference on proprietary infrastructure should expect that same pressure to reach its own roadmap eventually.

Cursor published the kernel on GitHub under the repository name mixture-of-kittens, along with the benchmark code used in its layer-level tests, and said it plans to keep maintaining the project and accept outside contributions. The announcement did not specify a software license for the release.

Teams training or serving mixture-of-experts models such as DeepSeek, GLM or Qwen on NVL72-class hardware now have a free, deterministic alternative to DeepEP and NCCL-based pipelines worth benchmarking against their current setup before locking in next year’s training infrastructure budget.

Cursor published this account of Mixture-of-Kittens, credited to researchers Stuart Sul, Nash Brown, Henry Wildermuth, William Lin and Federico Cassano, on its engineering blog in August 2026.