Moonshot AI
Kimi K2.5 benchmark results
13 results on 9 benchmarks from 3 sources.
Also listed as: kimi-k2.5-thinking · kimi-k2.5-instant · fireworks/kimi-k2p5 · Kimi K2.5 (Fireworks) · kimi-k2.5 · Kimi K2.5 (high)
All benchmark results
| Benchmarks | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|
| LMArena Text | 1450 | 78 | Overall, style control on; 74811 votes; 95% CI 1447–1453“kimi-k2.5-thinking” | Official leaderboard | Oct 8, 2026 |
| LMArena Text | 1430 | 113 | Overall, style control on; 8378 votes; 95% CI 1424–1437“kimi-k2.5-instant” | Official leaderboard | Oct 8, 2026 |
| LMArena WebDev | 1437 | 77 | Overall; 20127 votes; 95% CI 1431–1442“kimi-k2.5-thinking” | Official leaderboard | Oct 8, 2026 |
| LMArena WebDev | 1404 | 87 | Overall; 3269 votes; 95% CI 1392–1416“kimi-k2.5-instant” | Official leaderboard | Oct 8, 2026 |
| LMArena Vision | 1252 | 48 | Overall, style control on; 28326 votes; 95% CI 1246–1258“kimi-k2.5-thinking” | Official leaderboard | Oct 2, 2026 |
| LMArena Vision | 1237 | 62 | Overall, style control on; 4108 votes; 95% CI 1226–1248“kimi-k2.5-instant” | Official leaderboard | Oct 2, 2026 |
| GPQA Diamond | 87.6% | 68 | Epoch AI run, accuracy; effort: Fireworks | Independent run | Feb 2, 2026 |
| SWE-bench Verified | 73.8% | 17 | Epoch AI run (Epoch scaffold) | Independent run | Feb 17, 2026 |
| SWE-bench Verified | 70.8% | 25 | agent: mini-SWE-agent“kimi-k2.5” | Official leaderboard | Feb 17, 2026 |
| AIME (OTIS Mock AIME 2024–2025) | 92.2% | 64 | Epoch AI run, OTIS Mock AIME 2024–2025; effort: Fireworks | Independent run | Feb 2, 2026 |
| Humanity's Last Exam | 24.4% | 18 | Compiled by Epoch AI | Third-party compilation via Epoch AI | Jan 27, 2026 |
| OSWorld (Verified) | 63.3% | 2 | OSWorld; agent: Kimi-K2.5 | Official leaderboard via Epoch AI | Feb 3, 2026 |
| Terminal-Bench 2.0 | 43.2% | 84 | Terminal-Bench 2.0; agent: Terminus 2 | Official leaderboard via Epoch AI | Jan 27, 2026 |
FAQ
Where do these Kimi K2.5 scores come from?
From Epoch AI Benchmarking Hub, LMArena leaderboard dataset, and SWE-bench leaderboards (swebench.com). Each row links to the original leaderboard, dataset or report, and vendor-reported numbers are labelled.