Cursor just gave away the kernel it uses to train its own coding model, and the company says it delivers a 41% jump in training throughput on Nvidia's newest racks.
On August 4, Cursor Research released Mixture-of-Kittens, or MoK. It's a CUDA kernel that fuses the entire mixture-of-experts computation and network communication for a training step into a single, uninterrupted piece of code. Built for Nvidia's GB300 NVL72 systems, it runs up to 2.37 times faster than the strongest public baselines on a single MoE layer. End-to-end, it lifted training throughput at Cursor by 1.41x, from 760.9 tokens per second per GPU to 1,070.2, benchmarked across 512 GPUs spanning multiple NVL72 racks.
You don't need to know CUDA to get why this matters. Mixture-of-experts models like DeepSeek-V4, Qwen3.5, GLM-5.2 and Kimi K2.7 route each token through only a handful of specialised subnetworks instead of the whole model. That's what makes them cheaper to run than a dense model of the same size. There's a catch. The GPUs have to constantly shuttle tokens to whichever chip holds the right expert. Stitching that compute and communication together without stalling has been one of the hardest problems in training these models at scale. Cursor's own engineers say the MoE layer was eating more than half of end-to-end training time on Composer, its agentic coding model, before this release.
How the kernel actually works #
The kernel moves tokens between GPUs using pull-based dispatch and push-based combine. Ring token buffers batch the transfers, so the CPU never has to synchronise with the GPU mid-step. Then there's inter-SM overlapping: it splits a GB300's streaming multiprocessors so some are crunching matrix multiplications while others are still moving data. On Blackwell hardware, the kernel waits until it has exactly enough tokens - 256 of them - to trigger a full tensor-core matrix-multiply instruction. Then it fires while the next batch is still in transit.
None of that is new on its own. What's new is fusing all of it, computation and communication, into one kernel instead of stitching together separate libraries like NCCL, DeepEP or Megatron. That's what most labs still do.
MoK is also fully deterministic: the same training run produces the same result every time. That's harder to guarantee than it sounds, once you're overlapping communication and computation across 72 GPUs inside a single NVLink domain. The code is credited to Cursor engineers Stuart Sul, Nash Brown, Henry Wildermuth, William Lin and Federico Cassano. It's live now on GitHub, under Cursor's mixture-of-kittens repository. It follows an earlier post describing custom MXFP8 kernels that delivered a 1.5x training speedup on Blackwell. That project only tackled the compute side of the problem, and left communication as the next bottleneck.
Why a coding company is publishing chip benchmarks #
Frankly, the more interesting story here isn't the benchmark. It's who's publishing it. Cursor makes its money selling an AI coding assistant, not selling infrastructure. A GB200 NVL72 rack runs a hyperscaler roughly $3.9 million all in, once networking and storage are counted, according to SemiAnalysis. Nvidia has spent the past year selling racks like it on one promise: that software will keep unlocking more value from the same silicon. A company with no chip business, open-sourcing a kernel that squeezes 41% more throughput out of that exact hardware, is free evidence for Nvidia's pitch.
That's not charity.
Cursor trains Composer across tens of thousands of GPUs. Shaving training time isn't an academic exercise there. It's a line item. Every lab training a DeepSeek-style MoE model now has a free option for cutting the same bottleneck, at least on NVL72 hardware: Moonshot's Kimi, Zhipu's GLM, Alibaba's Qwen. Given how many frontier labs already run this architecture, that's not a small pool of potential users.
What happens next is less clear. Cursor hasn't said whether MoK will get backported to the older GB200 NVL72 racks still running in most data centres. Nor has it said whether an inference-side follow-up is coming, the way one did after the MXFP8 kernel post. For now, the code is public and the benchmarks are Cursor's own. The rest of the industry gets to check the math.
Also read: Blackstone Pitches Second Mega Debt Deal to Fund Anthropic's Google Chips • JPMorgan and Accenture Back a New Group to Standardize AI Token Costs • White House Flip-Flops on AI Regulation as California Law Takes Effect