Moonshot says Kimi K2.5 trained 1T parameters on 15.5T tokens without instability

CatAstro_Piyush · x · 2026-07-28

Moonshot’s Kimi K2.5 reportedly trained a 1T-parameter model on 15.5T tokens with no instability, thanks to Muon plus a simple QK clipping trick.

The attached graphic explains the failure mode: high-rank Muon updates can inflate QK dot products, which makes logits and loss blow up. A post-update rescale / QK clip bounds those dot products, preserving the Muon direction while keeping training stable.

Original post →

More from Models

Models channel →