Kimi K2 Base
Moonshot AI · 2025-07-11 · 1T parameters
Kimi K2 Base is the open-weight pre-trained foundation checkpoint of Moonshot AI's trillion-parameter Kimi K2 Mixture-of-Experts model, with about 32B activated parameters per token (384 experts, eight selected plus one shared). Released on July 11, 2025 alongside Kimi K2 Instruct for researchers who want to fine-tune it themselves, it was pre-trained on 15.5 trillion tokens with the MuonClip optimizer and has no instruction tuning or long-thinking mode. Its BF16 tensors contain 1,026,408,235,864 parameters and the configuration supports 131,072 positions via YaRN.
Benchmark scores
| Benchmark | Score |
|---|---|
| AGIEval English | 72.5 |
| BBH (BIG-Bench Hard) | 88.7 |
| C-Eval | 92.5 |
| Capability | 128.0 |
| Chinese SimpleQA | 77.6 |
| CMATH | 90.3 |
| CMMLU | 90.9 |
| CRUXEval-I (CoT) | 74.0 |
| CRUXEval-O (CoT) | 83.5 |
| EvalPlus (average) | 80.3 |
| GSM8K-Platinum | 94.2 |
| HumanEval | 78.2 |
| LiveCodeBench v6 - release-reported cohort | 26.3 |
| MBPP | 72.1 |
| MMLU-Redux | 90.2 |
| OpenBookQA | 50.8 |
| PIQA | 85.5 |
| RULER - 128K | 88.6 |
| RULER - 64K | 93.8 |
| SimpleQA | 35.3 |
| SuperGPQA | 44.7 |
| TriviaQA | 85.1 |