Overview
Kimi-Linear-48B-A3B-Base is the pretrained base checkpoint of Moonshot AI’s Kimi Linear language-model family. It has 48 billion total parameters, approximately 3 billion activated parameters, and a stated context length of 1 million tokens. The released checkpoints were trained on 5.7 trillion tokens. (github.com)
Capabilities and uses
- Generates text and provides a foundation for research, adaptation and fine-tuning, including long-context applications.
- Uses a hybrid architecture combining Kimi Delta Attention (KDA) and global Multi-Head Latent Attention (MLA) at a 3:1 ratio. Moonshot reports up to 75% less KV-cache use and up to 6× decoding throughput at 1 million tokens in its stated comparison; these are workload-dependent results, not guarantees. (github.com)
- Model weights are available from Moonshot’s model repository, which identifies the license as MIT. (huggingface.co)
Limitations
This is the Base checkpoint, not the separately released Instruct version; it may require fine-tuning before use as an instruction-following assistant. Generated text can be inaccurate, and long-context support does not guarantee accurate use of every token. Self-hosting requires suitable compute and compatible software. (github.com)
Release date
October 30, 2025. The Kimi Linear technical report was first submitted on that date; the supplied model record uses the same release date. (arxiv.org)
Pricing
Official input and output API prices for this specific Base checkpoint could not be verified in either CNY on the domestic platform or USD on the international platform. Prices for other Kimi models should not be applied to it. Checked November 17, 2026. The domestic pricing documentation was readable, while the international pricing documentation could not be retrieved; the international platform homepage was checked instead. (platform.kimi.com)