Prime Mega MoE makes a fast AI training kernel trainable
Prime Intellect released Prime Mega MoE, a backward pass for DeepSeek's fused Mega MoE kernel that makes the kernel trainable. The story says it speeds Qwen3 MoE training steps by up to 34% in prime-rl and gives a 3.11x faster forward pass. It is aimed at developers who train large AI models, not everyday users.
How to try it
Developers can check the prime-rl project to test it on Qwen3 MoE training, starting with the release post.
Useful 15%