Cursor Deploys 'MoK' to Boost GPU Training Throughput by 1.41x
Cursor has deployed a new training stack component called 'MoK' to power distributed training across tens of thousands of GPUs. This deployment has achieved a 1.41x increase in end-to-end training throughput compared to their previous DeepEP-based stack. Achieving a 1.41x throughput improvement at a scale of tens of thousands of GPUs is a significant milestone for machine learning infrastructure. It demonstrates that optimizing communication and training stacks can yield massive efficiency gains over state-of-the-art baselines like DeepEP. While specific technical details about MoK's architecture remain undisclosed, the performance gain was measured in a production environment. The baseline, DeepEP, is a highly optimized communication library designed by DeepSeek for Mixture-of-Experts (MoE) models.
## BACKGROUND
Large-scale AI training requires coordinating tens of thousands of GPUs, where communication overhead often becomes a major bottleneck. DeepEP (DeepEveryParallel) is an open-source, high-performance communication library developed by DeepSeek specifically for Mixture-of-Experts (MoE) models and expert parallelism. Optimizing these communication primitives is crucial for improving training speed and reducing infrastructure costs.