X · @teortaxesTex
· X / Twitter
Extremely interesting blog post on efficiently serving K3 They trained DSpark for a hybrid linear architecture! It's not even that much of a problem! …
Extremely interesting blog post on efficiently serving K3They trained DSpark for a hybrid linear architecture! It's not even that much of a problem! DSpark works everywhere, modulo skill issues.RadixArk: We trained a DSpark speculator draft model for Kimi K3 with SpecForge that takes batch-1 decode from ~113 to ~423 to