Cursor open-sources Mixture-of-Kittens, an MoE training megakernel

Cursor Research says it has open-sourced Mixture-of-Kittens (MoK), a production mixture-of-experts (MoE) training megakernel built for GB300 NVL72 systems.

What Cursor released

MoK combines MoE communication and computation into a single kernel and is designed to be fully deterministic. Cursor says the repository includes benchmark code for individual MoE layer benchmarks and is available at github.com/cursor/mixture-of-kittens.

The implementation supports BF16 and MXFP8 precision modes. Cursor says it uses MXFP8 for training while keeping the shared expert in BF16 for training stability.

Reported performance

In Cursor’s production training stack across several NVL72 racks, the company reports:

• 1.41× end-to-end tokens per second compared with its previous DeepEP-based setup
• An approximately 41% tokens-per-second speedup in that production comparison
• Up to 2.37× the fastest public baseline in the MXFP8 forward MoE-layer benchmark
• Up to 1.78× in the MXFP8 backward benchmark

These are Cursor’s own benchmark and production-stack results, not an independent comparison. The result depends on the workload, hardware, and baseline described in the post.

Why it matters

MoE systems spend a large part of training time moving tokens between experts. A kernel that coordinates communication and computation at the rack level can improve the efficiency of the training infrastructure behind large coding models, even though MoK is not an end-user Cursor editor feature.

For engineers working on distributed model training, the release is useful because the code, benchmark methodology, and hardware assumptions are public. For everyone else, the practical takeaway is that Cursor is publishing parts of the systems work that supports its agent and model development rather than keeping the entire stack proprietary.

Source

• Cursor Research — “Mixture-of-Kittens: MoE Megakernel for NVL72s” (August 2026): https://cursor.com/blog/mixture-of-kittens
• Official repository: https://github.com/cursor/mixture-of-kittens