Moonshot AIMoonlight

Moonlight-16B-A3B

Moonlight-16B-A3B is the base checkpoint from Moonshot's Muon training research. It uses a mixture of experts and was trained on 5.7 trillion tokens.

An experiment in training efficiency

The research adapts Muon to larger language models by adding weight decay and controlling the scale of parameter updates. It compares compute-efficient training against AdamW. This is a result about training, not a promise that every Moonlight inference will be twice as fast.Muon is Scalable for LLM Training: original report

Architecture compatibility and research artifacts

The release uses the DeepSeek-V3 implementation interface. That code compatibility does not establish that it was initialized from DeepSeek weights. Moonshot also supplies distributed Muon code and a report about intermediate training checkpoints. Those training artifacts belong to the same experiment.Moonlight research code and checkpoint information

Use the base checkpoint for adaptation

Practical guidance. Choose the base release when you need to study pretraining or train a task-specific adaptation. For a conversation interface, compare the separate Instruct checkpoint before adding your own instruction training.

Specifications & access

Moonlight-16B-A3B specifications
DetailValue and conditions
Launch dateFebruary 2025
AccessDownloadable
Model identifiersmoonshotai/Moonlight-16B-A3B
Context window8,192 tokens

Published checkpoint context limit; input and generated tokens share this budget.

Moonlight base checkpoint configuration
Parameters16 billion total; approximately 3 billion active per token.Moonlight-16B-A3B model card
Weight licenseMIT.Moonlight-16B-A3B model card
Training stagePretrained base model; the instruction-tuned release is separate.Moonlight-16B-A3B model card

Related models

  • Vision-language model for images, documents and video, built from Moonlight and the MoonViT encoder.

    Launch date April 2025Downloadable
  • Instruction-tuned Moonlight release for conversation and task prompts, with an 8,192-token context.

    Launch date February 2025Downloadable