Not an LLM-only trick. Train a Swin-B vision transformer from scratch on ImageNet-1K with Muon, and the same scheduler beats the best hand-tuned warm-up on top-1 accuracy. Same derivation, different…