Reasoning
198-question Diamond subset; independent AA evaluation
- 01
- 02
- 03
- 03
- 03
- 27
Kimi K2 Thinking
Across the task benchmarks we track, Kimi K2 Thinking’s strongest relative results are in coding, reasoning, and web research. On SWE-bench Verified, it scores 71.3%, ranking #6 among 7 models with reported results in our dataset.
Updated: Data as of:
Strongest results
| Benchmark | Measures | Score |
|---|---|---|
| SWE-bench ↗Verified / published open-model reports | Coding | 71.3%#6 of 7 |
| Humanity's Last Exam ↗Stable HLE — text-only subset | Expert reasoning | 23.82%#57 of 68 |
| IFBench ↗58 constraint types | Precise instruction following | 68.1%#15 of 17 |
| GPQA ↗Diamond / Artificial Analysis | Reasoning | 83.84%#27 of 30 |
Model details
| Model | Kimi K2 Thinking |
|---|---|
| Agent Index v1.2 | Insufficient coverage (0/7) · Preview |
| Creator | Moonshot AI (opens in a new tab) |
| Availability | Open source / open weights |
| Context size | 256,000 tokens (opens in a new tab) |
| Max output | Not reported |
| Knowledge cutoff | Not reported |
| License | Modified MIT License · Model weights ↗ (opens in a new tab) |
Pricing and performance
Per provider and reported configuration. Prices are USD per 1M tokens.
No provider pricing or performance reports have been added for this model yet.
Benchmark performance
Kimi K2 Thinking in color, alongside leading models. Latest stored reports, best effort per report date.
GPQA Diamond ↗
Coding
Verified split; published reports
- 01
- 02
- 03
- 04
- 05
- 06
Web Research
Published reports; source settings retained
- 01
- 02
- 03
- 04
- 05
- 08
Reported benchmarks
All published results in our dataset, with versions, evaluation settings, and sources.
| Benchmark | Score | Evaluation setup | Source |
|---|---|---|---|
| BrowseCompPublished open-model reportsAccuracy · Higher is better | 60.2% | harness comparison: provider-reported; settings differ across models · 256K context; 300-step limit; 24K reasoning per step; hide old tool outputs; Hugging Face blocked | Moonshot AI ↗ (opens in a new tab)provider reported |
| GPQADiamond / Artificial AnalysisAccuracy · Higher is better | 83.838% | Kimi K2 Thinking; 198-question Diamond subset; independent AA evaluation | Artificial Analysis (opens in a new tab)independent |
| Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better | 23.9% | Provider no-tools text-only evaluation. Distinct from independent AA 23.8. · subset: text · No tools | Moonshot AI ↗ (opens in a new tab)provider reported |
| Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better | 51% | Heavy configuration: parallel test-time compute. Text-only; retained separately from single-agent 44.9. · subset: text · Tools enabled | Moonshot AI ↗ (opens in a new tab)provider reported |
| Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better | 44.9% | Text-only; 120 steps; 48K reasoning per step; 256K context; o3-mini judge; Hugging Face blocked · Tools enabled | Moonshot AI ↗ (opens in a new tab)provider reported |
| Humanity's Last ExamStable HLE — text-only subsetAccuracy · Higher is better | 23.818% | Kimi K2 Thinking; original HLE text-only 2,158 questions; no tools; not HLE Diamond · No tools | Artificial Analysis (opens in a new tab)independent |
| IFBench58 constraint typesAccuracy · Higher is better | 68.095% | selection: best reported configuration per release · Kimi K2 Thinking · tools: benchmark-defined | Artificial Analysis (opens in a new tab)independent |
| SWE-benchVerified / published open-model reportsResolved · Higher is better | 71.3% | harness comparison: provider-reported; settings differ across models · Moonshot in-house SWE-agent-derived harness; clamped Bash/Edit context and tailored prompt | Moonshot AI ↗ (opens in a new tab)provider reported |