Prime Intellect's New MoE Kernel Speeds Qwen3 Model Training by Up to 55%

Prime Intellect released Prime Mega MoE, a tool that optimizes training of large-scale Mixture-of-Experts AI models, reportedly accelerating Qwen3-235B-A22B training by 55% and Qwen3-30B-A3B by 48% compared to TorchTitan.