According to CoinMeta, the mixture-of-kittens ( mok ) open-sourced by Cursor is used to accelerate the training of large MoE models. This technology combines the originally separate data transmission and computing processes into one, allowing GPU to perform computations while transmitting data, thereby reducing waiting times. In actual training with 512 GB300 GPU, the overall throughput increased by 41%. When testing the MoE layer separately, the speed of mok was up to 2.37 times faster than that of public solutions. mok has already been used in the training of tens of thousands of GPU with Cursor and is open-sourced under the Apache 2.0 license. Currently, it only supports NVIDIA Blackwell GPU and is mainly aimed at institutions that possess GB200 and GB300 NVL72 clusters.