Prime Intellect Releases Prime
Prime Intellect keeps quietly building the plumbing for decentralized training, and this is a solid piece of it. The clever part is not the 3.11x forward speedup — kernel vendors demo those all the time — but the backward pass, since a fused kernel you cannot differentiate is just an inference toy.… more
Making DeepSeek's Mega MoE trainable and cutting Qwen3 MoE step time by up to a third in prime-rl directly serves the 'train big models outside the hyperscalers' thesis. In our view it is a genuine engineering win rather than hype, though the numbers are self-reported and will vary across cluster setups. If you follow the decentralized-training scene, this is one of the more concrete datapoints lately.
Prime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass
Prime Intellect released Prime Mega MoE, a backward pass for DeepSeek’s fused Mega MoE kernel. It makes the kernel trainable and speeds Qwen3 MoE training steps by up to 34% in prime-rl. The post Prime Intellect Releases Prime Mega MoE: Trainable Fused M…
Source:
Marktechpost
Discussion (0)