Overview
Kimi-Linear-48B-A3B-Instruct is an instruction-tuned, text-generation model designed in part for efficient long-context processing. It has 48 billion total parameters, approximately 3 billion activated parameters, and a stated context length of up to 1 million tokens. Its hybrid architecture combines Kimi Delta Attention (KDA) and Multi-Head Latent Attention (MLA) at a 3:1 ratio. (huggingface.co)
Capabilities and uses
The model can be used for instruction-following, conversational text generation, and tasks involving long documents or other long text contexts. Moonshot reports up to 75% lower KV-cache usage and up to six times the decoding throughput at a 1-million-token context compared with its full-MLA baseline; these are results under the reported experimental conditions, not guarantees for every deployment. (huggingface.co)
Availability and limitations
Moonshot has released the model checkpoint; the official model card lists the MIT license. The card describes deployment with Transformers or vLLM. The maximum usable context and performance depend on the deployment environment and available compute. Outputs can be incorrect and should be checked, particularly for consequential uses. (huggingface.co)
Release date
October 30, 2025. The official model repository was created on that date, and the Kimi Linear technical report was first submitted the same day. (huggingface.co)
Pricing
Official input and output API pricing for this exact model could not be verified in either CNY on the domestic platform or USD on the international platform. The official pricing pages were checked on November 17, 2026; their accessible text did not provide a model-specific quote. Prices for other Kimi models should not be applied to this model. (platform.kimi.com)