Kimi Linear 48B A3B Base
Moonshot AI · 2025-10-30 · 49.1B parameters
Kimi Linear 48B A3B Base is the pre-trained base checkpoint of Moonshot AI's Kimi Linear hybrid linear-attention MoE model, released October 30, 2025 with 48B total and 3B active parameters. It interleaves Kimi Delta Attention (KDA) layers with global MLA layers at a 3:1 ratio, which cuts KV-cache use by up to 75 percent, and was trained on 5.7 trillion tokens; the card lists a 1M-token context. Its tensors contain 49,122,681,728 parameters.
Benchmark scores
| Benchmark | Score |
|---|---|
| C-Eval | 83.3 |
| Capability | 117.6 |
| Chinese SimpleQA | 53.5 |
| CMATH | 85.5 |
| CRUXEval-I (CoT) | 61.0 |
| CRUXEval-O (CoT) | 67.0 |
| EvalPlus (average) | 64.9 |
| GPQA Diamond | 40.4 |
| GSM8K | 86.3 |
| GSM8K-Platinum | 89.6 |
| LiveCodeBench v6 - release-reported cohort | 20.0 |
| MATH | 58.5 |
| MMLU Pro | 54.8 |
| MMLU-Redux | 79.7 |
| SimpleQA | 10.1 |
| TriviaQA | 75.2 |
| WinoGrande | 81.5 |