Kimi K2: benchmark results
Moonshot Kimi K2 open 1T-parameter MoE model (32B active). Provider: Moonshot. Released 2025-07-11. Access: Open.
Aggregate Rating 1681 ± 3, rank #243 of 1656 rated models, from 1817 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentHazard (Claude Code) - Compliance Bypass | 20.62 | Attack success rate (%) on the compliance bypass instances ( | 100 |
| CLEM Wordle | 73 | Game Clemscore (%), text v3.0; published played × quality / | 100 |
| ControBench - Religion | 53.71 | Macro-F1 (%; 9-class user stance, one-shot prompting on a st | 100 |
| ControBench - Trump | 89.46 | Macro-F1 (%; 2-class user stance, one-shot prompting on a st | 100 |
| Ebisu - JF-TE - F1 | 43.96 | Maximal-Term F1 (%) | 100 |
| FinReasoning (Deep Insight) | 91.6 | Rubric score (0-100): a judge score fused from DeepSeek-V3 a | 100 |
| FinReasoning (Deep Insight) - Completeness and Comparative Analysis | 89.5 | Rubric score (0-100): a judge score fused from DeepSeek-V3 a | 100 |
| FinReasoning (Deep Insight) - Justification and Causal Depth | 96.3 | Rubric score (0-100): a judge score fused from DeepSeek-V3 a | 100 |
| KOCO-BENCH | 8.9 | Average Pass@1 (%) | 100 |
| LingxiDiagBench (Dynamic, Free-Form) - 4-Class Accuracy | 36 | Accuracy (%, 0-100) on four-way classification (pure depress | 100 |
| LogiScope-VQA - Warehouse Commonsense | 96 | Micro-averaged accuracy (%) | 100 |
| MLB | 77.29 | Average (self-reported) | 100 |
Interactive version: theaggregate.ai/model?slug=kimi-k2 · How It Works · Data refreshed daily, snapshot 2026-10-20.