# Cursor open-sources MoK after a claimed 41% training throughput gain > Source: > Published: 2026-08-04 17:27:08+00:00 [Cursor (@cursor_ai)](https://x.com/cursor_ai/status/2084670806613737919?s=46) open-sourced Mixture-of-Kittens on August 4th, a GPU kernel that the AI coding company says increased end-to-end throughput for training its Composer models by 41% against its previous production stack. The release exposes a piece of the model infrastructure that CEO [Michael Truell (@mntruell)](https://x.com/mntruell) and Cursor's research operation have built beneath the company's coding editor. Truell co-founded Cursor's parent, Anysphere, with three MIT classmates in 2022. Cursor has since moved deeper into model development, introducing its [Composer 2 coding model](https://cursor.com/blog/composer-2) in March after what it described as its first continued pretraining run. [Mixture-of-Kittens](https://cursor.com/blog/mixture-of-kittens), or MoK, targets one of the costliest parts of training mixture-of-experts models: moving tokens among the GPUs that hold different experts, running the selected feed-forward networks and returning the results. Cursor's researchers found that this layer could consume more than half of total training time, depending on the workload and configuration. MoK combines those communication and computation steps inside a single megakernel designed for Nvidia GB300 NVL72 systems. An NVL72 rack links 72 GPUs in one NVLink domain, creating opportunities to overlap data transfers with computation. It also forces infrastructure teams to carefully coordinate work across the entire rack rather than optimizing each matrix multiplication in isolation. ### Cursor's benchmark claims Cursor says MoK delivered up to 2.37x the MXFP8 forward throughput of the fastest public baseline in its layer-level tests. The comparisons covered NCCL with PyTorch, DeepEP with PyTorch, DeepEP with Transformer Engine, and Nvidia's HybridEP with Megatron setup. The peak figure is one result within Cursor's own tests, rather than a uniform improvement across every precision mode and training pass. Cursor reported maximum gains of 1.78x for MXFP8 backward execution, 1.92x for BF16 forward execution and 1.58x for BF16 backward execution. For the production comparison, Cursor tested 512 GPUs distributed across several GB300 NVL72 racks. Its previous DeepEP-based stack processed 760.9 tokens per second per GPU, while MoK processed 1,070.2, according to the company. That produced the reported 1.41x end-to-end increase. The production number comes from Cursor's internal training system and cannot be reproduced from the public repository alone. Cursor has, however, released the code used for its individual MoE-layer benchmarks, giving infrastructure engineers a way to test the narrower performance claims on compatible hardware. Cursor says MoK already powers Composer training across tens of thousands of GPUs. That deployment figure and the benchmark results are company-reported. ### Rebuilding the MoE layer around communication The implementation uses pull-based dispatch, where a GPU retrieves the tokens it needs from peers, followed by push-based combine operations. Cursor says that arrangement reduces synchronization among the 72 GPUs and delivered up to 29% higher NVLink bandwidth utilization than push-based dispatch under uneven expert loads. MoK divides GPU streaming multiprocessors between communication and computation, then adjusts how many tokens move through the system in each minibatch. Ring token buffers let the kernel reuse a fixed memory allocation without waiting for the CPU to calculate and allocate buffers for each changing token count. The kernel also fixes the order of floating-point operations, producing bitwise-identical outputs for the same inputs. Cursor said it needed that determinism for internal experiments and on-policy reinforcement-learning post-training. MoK supports BF16 and MXFP8, with Cursor using MXFP8 in production while keeping the shared expert in BF16 for training stability. The [GitHub repository](https://github.com/cursor/mixture-of-kittens) is available under the Apache 2.0 license. The named authors are Stuart H. Sul, Nash Brown, Henry Wildermuth, William Lin and Federico Cassano. ### Cursor moves further down the model stack Cursor's release makes visible how much infrastructure sits beneath an AI application that began as a code editor. The company is training proprietary coding models, operating them across large GPU clusters and writing hardware-specific kernels when public implementations leave too much performance unused. That work is backed by a large capital base. In its November 13th, 2025 [Series D announcement](https://cursor.com/blog/series-d), Cursor said it raised $2.3 billion at a $29.3 billion post-money valuation from investors including Accel, Thrive Capital, Andreessen Horowitz, Coatue, Nvidia and Google. Cursor said at the time that the financing would support deeper research investment. Open-sourcing MoK gives other labs access to a production-tested implementation while inviting outside engineers to inspect and extend it. The immediate audience remains narrow: teams training large mixture-of-experts models on GB300 NVL72 systems. Those are costly machines, and the software is tailored to their networking and execution model. The release also gives Cursor a public artifact for recruiting researchers who might otherwise associate the company mainly with its editor. Cursor says AI agents helped its small research group write individual kernels and work through the distributed megakernel. MoK is the resulting infrastructure bet: better coding models will depend partly on making each training cluster deliver more useful work per second.