Our hope is that this lowers the barrier to AI research, so more labs can train

Cursor’s Mixture-of-Kittens targets a bottleneck in training models whose specialized components are spread across graphics processors. Data must travel to the selected components, be processed and return. MoK overlaps that movement with computation inside a combined GPU program, reducing time spent waiting between operations. In its 512-GPU production-stack comparison, Cursor reports about 41% higher training throughput on GB300 NVL72 hardware than its previous DeepEP-based setup.

The larger advertised gain of up to 2.37 times applies to one layer’s MXFP8 forward computation against the fastest public baseline. It is a different measurement from the complete-training result, so it should not be read as training an entire model 2.37 times faster.

“As we have scaled the training and inference of Composer , our agentic coding model, the mixture-of-experts layer has consistently remained the major bottleneck. Depending on the workload and training configuration, it can consume more than half of end-to-end training time.”

cursor_ai

“MoK addresses that bottleneck by fusing all MoE communication and computation into a single, fully deterministic kernel. It now powers Composer training across tens of thousands of GPUs.”

cursor_ai

“Mixture-of-Kittens achieves up to 2.37x higher MXFP8 forward throughput than the fastest public baseline on GB300 NVL72s”

cursor_ai

“In our production training stack across several NVL72 racks, MoK increased end-to-end tokens per second by 1.41x.”

cursor_ai

“Empirically, pull-based communication delivers up to 29% higher NVLink bandwidth utilization than push-based communication when dispatching tokens under expert imbalance, making it the better choice.”

cursor_ai