/TECHStory update pending

MiMo V2.5 was trained on 48 trillion multimodal tokens, says research engineer Elie Bakouch

The text-only pro model was trained on 27 trillion tokens

Story Brief

The text-only pro model was trained on 27 trillion tokens

Source Posts on X

Cluster sources

@teortaxesTex yea true, just checked it's trained on 48T while pro is train on 27T (text only)

@eliebakouch but so is MiMo V2.5.