WRSWorld Research StateContribute
WORLD RESEARCH STATE

Moonlight: update scaling and distributed Muon training

Contribution grade

Large-language-model optimization systems · Human-designed research method or resource

Reference score70,000Provisional editorial estimate of this exact contribution; its grade is assessed separately.

What changed

Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments.

Scientific scope

Added weight decay and adjusted parameter-update scales for large-model Muon, implementing ZeRO-style distributed orthogonalization and releasing Moonlight checkpoints with scaling experiments. Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Extends rather than invents Muon/NewtonSchulz/ZeRO, and differs from KimiK2 MuonClip. Approximately twofold compute efficiency is fitted/reported for studied training regimes, not universal wall-clock superiority or tuning-free performance for every architecture.

Contributors

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction