Stable LatentMoE：KDA+LatentMoE+AttnRes三合一，训练成本接近K2

> Stable LatentMoE so we have at least three parts: KDA, LatentMoE, AttnRes yeah, a thing of beauty and the funniest part: it didn't cost them that much more than K2 to train If they throw a few billions at compute, Fable-Schmable becomes a bullet point rather than a milestone