← All models

Moonlight-16B-A3B

Published Publisher Moonshot AI

Model positioning

Moonlight-16B-A3B is a pretrained base language model released by Moonshot AI. It uses a mixture-of-experts (MoE) architecture, with approximately 16 billion total parameters and approximately 3 billion parameters activated per token. It was trained on 5.7 trillion tokens using the Muon optimizer. This base model differs from the separately released instruction-tuned version, Moonlight-16B-A3B-Instruct. (huggingface.co)

Capabilities and uses

  • Supports text generation, with an officially stated context length of 8K tokens.
  • Official evaluations cover English knowledge, Chinese, coding, and mathematics tasks; for example, the reported MMLU score is 70.0 and HumanEval score is 48.1. Evaluation results do not guarantee performance in specific applications. (huggingface.co)
  • Can be used for research on language models and training optimization, text generation experiments, and self-deployment. The official project provides a Transformers usage example and states that inference engines such as vLLM and SGLang can be used. (github.com)
  • The model weights have been officially released, and the model card specifies an MIT license. (huggingface.co)

Limitations

This is a pretrained base model and should not be equated with the instruction-tuned version. 8K is the officially listed context length and does not represent a separate maximum output length; the knowledge cutoff date and separate maximum output length have not been verified. Self-deployment still incurs computing, storage, and operational costs.

Release date

2025-02-22; the official model repository's commit history shows an initial commit on that date. (huggingface.co)

Pricing

As of 2026-11-19, checks of the official domestic and international pricing pages did not verify official hosted API input or output prices for this model; therefore, no RMB or USD amounts, billing units, or pricing tiers can be confirmed. Prices for other Kimi models are not applied to this model. (platform.kimi.com)

Official sources