Model Overview
Moonlight-16B-A3B-Instruct is an instruction-fine-tuned text generation 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. Moonlight uses the Muon optimizer and was trained on 5.7 trillion tokens of data. (huggingface.co)
Capabilities and Uses
The model is designed for instruction following and conversational text generation. It can be used for question answering, summarization, writing assistance, and experiments with coding and mathematical tasks. The official model card provides benchmark evaluations related to English, Chinese, coding, and mathematics; these results should not be regarded as a guarantee of performance in specific applications. (huggingface.co)
The model weights have been officially released, and the model page specifies an MIT license; the model can be used for local or self-hosted inference and research. Deployers remain responsible for computing resources and operational costs and must comply with the license terms. (huggingface.co)
Limitations
max_position_embeddingsin the official configuration is 8,192; long-document processing is constrained by the context length, which should not be treated as a separate maximum output length. (huggingface.co)- The model is primarily intended for text generation; native image, audio, or video processing capabilities have not been verified.
- Generated content may be inaccurate; users should verify code and important facts. Resources required for self-hosting depend on precision, inference framework, and deployment configuration.
- The model ID for the official hosted API has not been verified; the Hugging Face repository identifier should not automatically be treated as the official API ID.
Release Date
February 22, 2025. The official model repository's commit history shows an initial commit and model file commits on that date; this date is based on repository records and does not imply a separate release announcement. (huggingface.co)
Pricing
As of November 19, 2026, the official domestic and international pricing pages had been consulted, but official hosted API input and output prices for this model had not been verified, nor had the applicable currency or billing units; therefore, prices for other Kimi models or third-party services are not applied to this model. Hardware and operating costs for self-hosting are additional. (platform.kimi.com)