Prime Intellect released Prime Mega MoE — the backward pass for the fused Mega MoE kernel from DeepSeek. According to MarkTechPost, this makes the kernel trainable.
In the publication's title, forward-pass acceleration is stated as 3,11 times, and in the description — the acceleration of training steps for Qwen3 MoE in prime-rl to 34%. These figures relate to different formulations and require clarification of the methodology.
Practical meaning of the news is a potential reduction in training time for mixture-of-experts models. However, the available information is presented as metadata of an independent publication, not as full technical documentation or confirmation from the primary source.
editorial commentary
Why it matters
Probable consequence — developer interest in using Prime Mega MoE in training. The next observed signal — publication of technical tests, code, or confirmation from the primary source. Substantial uncertainty is tied to the absence of a complete methodology and independent verification in the package.