Frontier Pre-Training at Microsoft AI

psoulos 40 points 4 comments October 15, 2026
microsoft.ai · View on Hacker News

Discussion Highlights (3 comments)

Bolwin

Very similar to Kimi K3 it seems, but fewer active params (80b vs 104b). I don't think they mentioned what form of attention they use

adroitboss

Wait, I don't get it. Is the model not open? If not, why are they comparing it to Open Weight models instead of the state of the art?

dsign

Microsoft should stop tormenting developers and fix text input in MS Teams, then go and atone for its many, many sins, before committing new ones, like this “this not really an open weights model but we want to IE6-them with our bully tactics”.

Semantic search powered by Rivestack pgvector
9,619 stories · 90,738 chunks indexed