AIMarkTechPostabout 3 hours ago
Prime Intellect Releases Prime Mega MoE Kernel
4 min read
Prime Intellect released Prime Mega MoE, a training extension for DeepSeek's fused Mega MoE kernel, enabling fast overlapping backward passes. On Qwen3 MoE models, full training steps run up to 34% faster than the prime-rl baseline and 55% faster against TorchTitan. The kernel runs on NVIDIA Hopper and Blackwell GPUs using BF16, MXFP8, or blockwise FP8 precisions.
Level
Hype check
Tap to vote and see what everyone thinks.
