Matched controlled run on FineWeb-Edu, 124M model, BS=24, compile on, seed 1234.
Muon (Muon on 2D weights + AdamW on embeddings/norms) final loss 3.557 vs
AdamW 3.724 at 300M tokens (-0.167); reaches AdamW's final loss at ~165M tokens
(~1.8x token-efficiency); MFU cost ~0% (full-run median). 0 NaN skips both.
Artifacts: ABLATIONS.md, ablations.json, per-variant metrics.jsonl, run logs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>