Kimi-Linear-48B-A3B-Instruct
Kimi Linear Instruct applies instruction tuning to the 48-billion-parameter Kimi Linear model. It retains the long context and sparse expert design of the base release.
The speed result has a specific workload
Moonshot reports up to a 75% reduction in attention-cache use and roughly sixfold decoding throughput at a one-million-token context compared with full MLA in its experiments. These are workload-specific comparisons, not a fixed speed increase for every prompt.Kimi Linear: original attention architecture report
Use the instruction checkpoint and its template
The model card provides a chat-template example and a vLLM configuration with a 1,048,576-token limit. It also describes use through Transformers with the required linear-attention dependencies. Downloading weights alone does not supply those kernels.Kimi-Linear-48B-A3B-Instruct model card
Measure the parts of response time separately
Practical guidance. Compare prompt processing and token generation separately. A long-context decoding improvement can matter greatly for one workload and have little effect on a short prompt with a brief answer.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | October 2025 |
| Access | Downloadable |
| Model identifiers | moonshotai/Kimi-Linear-48B-A3B-Instruct |
| Context window | 1,048,576 tokens Published checkpoint context limit; input and generated tokens share this budget. Kimi-Linear-48B-A3B-Instruct model card |
| Parameters | 48 billion total; 3 billion active per token.Kimi-Linear-48B-A3B-Instruct model card |
| Weight license | MIT.Kimi-Linear-48B-A3B-Instruct model card |
| Serving requirements | Published examples use the Kimi Linear implementation and attention kernels.Kimi-Linear-48B-A3B-Instruct model card |
Related models
- Kimi K3.1 Pro
Moonshot AI · Kimi K3.1
A K3-derived multimodal model with sparse attention, 2.8T total parameters and 104B active parameters.
Launch date not confirmedAPI · Downloadable - Kimi-Linear-48B-A3B-Base
Moonshot AI · Kimi Linear
Pretrained hybrid-attention model that combines recurrent memory with selected global-attention layers.
Launch date October 2025Downloadable