Up to four models
Kimi K3 vs Kimi K2.6
On the index (GPT-4 = 100), Kimi K3 scores 168 and Kimi K2.6 152. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
Kimi K3
168152–183
WeightsOpen
ReleasedJul 2026
Price in / out$0.19 / $16
Context1.05M
Core tests taken6 of 9
Kimi K2.6
152136–169
WeightsOpen
ReleasedApr 2026
Price in / out$0.65 / $3.41
Context262K
Core tests taken4 of 9
Kimi K3 leads Kimi K2.6 by 15 points, ahead on 3 of 3 shared core tests, at 3.1× the price.
Test by testthird-party results
Kimi K3Kimi K2.6
Every result as a table
| Test | Kimi K3 | Kimi K2.6 |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 93.1% | 90.8% |
| LiveBench Data Analysis | 78.7% | 65.1% |
| LiveBench IF | 71.4% | 64.4% |
| LiveBench Language | 85.5% | 75.1% |
| LiveBench Reasoning | 90.7% | 79.4% |
| GPQA Diamond (Vals AI) | 92.9% | 89.1% |
| MMLU Pro | 88.0% | 87.6% |
| MMMU Pro | 88.1% | 86.3% |
| SimpleQA Verified | 50.6% | 34.9% |
| Maths | ||
| FrontierMath (tiers 1-3) | 72.2% | 57.2% |
| AIME-style maths (OTIS mock) | 97.2% | 96.1% |
| LiveBench Mathematics | 84.4% | 84.3% |
| Chess puzzles | 37.0% | 20.0% |
| FrontierMath (tier 4) | 39.0% | 25.6% |
| Coding | ||
| LiveBench Agentic Coding | 62.2% | 46.9% |
| LiveBench Coding | 81.5% | 78.6% |
| Code Migration | 16.1% | 27.8% |
| LiveCodeBench | 87.2% | 86.8% |
| ProgramBench | 2.0% | 0.0% |
| SWE-bench | 93.4% | 76.2% |
| Vibe Code Bench v1.1 | 85.0% | 37.9% |
| WeirdML | 82.6% | 55.9% |
| Agents | ||
| DTBench | 91.2% | 90.9% |
| Furniture assembly | 34.2% | 21.7% |
| LMCA | 52.7% | 37.3% |
| Mystery game puzzles | 26.0% | 18.0% |
| Professional work | ||
| Finance Agent (v2) | 53.1% | 44.9% |
| CorpFin v2 | 71.6% | 66.7% |
| EMB | 66.7% | 57.9% |
| LegalBench | 86.2% | 84.7% |
| Legal Research Bench | 46.2% | 15.9% |
| MedCode | 49.4% | 40.1% |
| MedScribe | 88.0% | 78.1% |
| MortgageTax | 66.3% | 65.8% |
| Public Benefits Bench v1.1 | 68.2% | 56.6% |
| Tax Agent Bench | 33.4% | 12.4% |
| TaxEval v2 | 75.7% | 74.7% |
Tests at least two of these models have taken. Bold rows are the core tests.