# MoonshotAI/Moonlight

Muon is Scalable for LLM Training

Repository: https://github.com/MoonshotAI/Moonlight
Canonical: https://ross.abutalabs.com/products/moonlight
License: MIT
License Family: permissive
Last push: 2025-08-03T06:31:18+00:00

## Health v2 (maintenance only)
Score: 36/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 35, release rhythm 35, longevity 39
- inputs: {"age_days": 557, "days_push": 395, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1540, forks 101 (observed 2026-08-28T04:05:00.467658+00:00)

## What it is
Moonlight is a 3B/16B Mixture-of-Experts LLM trained with the Muon optimizer, released by Moonshot AI along with a memory- and communication-efficient distributed Muon implementation. It includes pretrained, instruction-tuned, and intermediate checkpoints plus scaling law experiments showing ~2x compute efficiency over AdamW.

## Use cases
- train large language models with the Muon optimizer
- run distributed Muon training with ZeRO-1 memory optimization
- compare Muon vs AdamW scaling law experiments
- fine-tune or study an open MoE model checkpoint
- reproduce compute-optimal LLM training research

## When to choose
- you want to experiment with Muon for large-scale LLM training
- you need an efficient distributed optimizer implementation for research
- you want open MoE model checkpoints for study or fine-tuning

## When to avoid
- you need a production-ready training framework with broad ecosystem support
- you lack multi-GPU infrastructure for large model training
- you only need inference rather than training

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, deep-learning
- domain: large-language-models, deep-learning, machine-learning
- platform: python
- tags: muon-optimizer, mixture-of-experts, scaling-laws, distributed-training, model-checkpoints, gpu, linux

## Member repositories
- MoonshotAI/Moonlight (main) score 36

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:00.467658+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:30:49.838437+00:00, confidence not recorded.
  - readme: https://github.com/MoonshotAI/Moonlight (fetched 2026-08-28T04:05:00.467658+00:00, sha daebb2e344ca)
- Data as of 2026-08-30T08:39:29.467469+00:00.