# Kimi K2.5 vs Kimi K3

> Side-by-side benchmark comparison for Kimi K2.5 and Kimi K3 across agentic, coding, multimodal, knowledge, reasoning, multilingual, and math tasks.

- Canonical page: https://benchlm.ai/compare/kimi-k2-5-vs-kimi-k3
- Last updated: October 2, 2026

- Shared sourced benchmarks: 8
- HTML indexing: indexable
- Ranking lane: BenchAlign v5.8

## Quick Verdict

Pick Kimi K3 if you want the stronger benchmark profile. Kimi K2.5 only makes more sense when its price, context window, or workload-specific category wins matter more than the overall score.

## Summary

- Kimi K3 leads overall 72.14 to 53.67.
- The clearest category separation is in agentic, where the averages are 38 for Kimi K2.5 and 68.1 for Kimi K3.
- The biggest single benchmark swing is MCP Atlas in Agentic, with scores of 29.5% and 84.2%.
- Kimi K2.5 is the cheaper option on output tokens, which matters if you expect large responses or heavy interactive use.
- Kimi K3 also has the larger context window at 1.05M.

## Model Snapshot

| Property | Kimi K2.5 | Kimi K3 |
|----------|----------|----------|
| Creator | Moonshot AI | Moonshot AI |
| Type | Open Weight | Pending |
| Reasoning | Non-Reasoning | Reasoning |
| Context | 256K | 1.05M |
| Overall Score | 53.67 | 72.14 |
| Benchmarks Covered | 45 | 48 |
| Pricing (input/output) | $0.60 / $3.00 | $3.00 / $15.00 |

## Category Breakdown

### Agentic

- Winner: Directional only
- Kimi K2.5 public-lane score: 38 (Estimated · #63/119)
- Kimi K3 public-lane score: 68.1 (Supported · #8/119)

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| Terminal-Bench 2.0 | 50.8% | Coming soon | Coming soon |
| BrowseComp | 60.6% | 91.2% | Kimi K3 |
| Claw-Eval | 52.3% | Coming soon | Coming soon |
| QwenClawBench | 54.3% | Coming soon | Coming soon |
| τ³-bench results | 65.7% | Coming soon | Coming soon |
| DeepSearchQA | 77.1% | 95.0% | Kimi K3 |
| DeepPlanning | 14.4% | Coming soon | Coming soon |
| Toolathlon | 27.8% | Coming soon | Coming soon |
| MCP Atlas | 29.5% | 84.2% | Kimi K3 |
| MCP-Tasks | 59.1% | Coming soon | Coming soon |
| WideResearch | 72.7% | Coming soon | Coming soon |
| Gert Labs | 45.88% | Coming soon | Coming soon |
| ResearchClawBench | 14.0% | Coming soon | Coming soon |
| JobBench | 8.7% | 52.9% | Kimi K3 |
| Terminal-Bench 2.1 | Coming soon | 88.3% | Coming soon |
| Toolathlon-Verified | Coming soon | 73.2% | Coming soon |
| AutomationBench | Coming soon | 30.8% | Coming soon |
| APEX-Agents | Coming soon | 37.6% | Coming soon |
| SpreadsheetBench 2 | Coming soon | 34.8% | Coming soon |
| DECK-Bench | Coming soon | 73.5% | Coming soon |
| Terminal-Bench 2.1 (Vals) | Coming soon | 80.9% | Coming soon |
| ApprenticeBench | Coming soon | 18% | Coming soon |

### Coding

- Winner: Directional only
- Kimi K2.5 public-lane score: 40.1 (Estimated · #64/144)
- Kimi K3 public-lane score: 61.4 (Supported · #18/144)

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| SWE-bench Verified | 76.8% | Coming soon | Coming soon |
| SWE-bench Verified* | 70.8% | Coming soon | Coming soon |
| LiveCodeBench v6 | 85.0% | Coming soon | Coming soon |
| SWE-bench Pro | 50.7% | Coming soon | Coming soon |
| SWE Multilingual | 73% | Coming soon | Coming soon |
| SWE-Rebench | 58.5% | Coming soon | Coming soon |
| React Native Evals | 77.2% | Coming soon | Coming soon |
| SciCode | 48.7% | Coming soon | Coming soon |
| DeepSWE | Coming soon | 67.5% | Coming soon |
| CursorBench 3.2 | Coming soon | 60.8% | Coming soon |
| FrontierSWE | Coming soon | 81.2% | Coming soon |
| ProgramBench | Coming soon | 77.8% | Coming soon |
| Kimi Code Bench v2 | Coming soon | 72.9% | Coming soon |
| sweMarathon | Coming soon | 42% | Coming soon |
| PostTrain Bench | Coming soon | 36.6% | Coming soon |
| MLS-Bench Lite | Coming soon | 48.3% | Coming soon |
| VulcanBench v3 | Coming soon | 73.7% | Coming soon |
| OpenHarmony Bench | Coming soon | 57.3% | Coming soon |
| FrontierSWE v2 | Coming soon | 25.9% | Coming soon |
| LiveCodeBench (Vals) | Coming soon | 87.2% | Coming soon |
| SWE-bench (Vals) | Coming soon | 93.4% | Coming soon |
| PostTrainBench v1.1 | Coming soon | 32.0% | Coming soon |

### Reasoning

- Winner: Not comparable
- Kimi K2.5 public-lane score: 55.2 (Unranked · 3 rankable rows)
- Kimi K3 public-lane score: 65.8 (#18/27)

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| LongBench v2 | 61% | Coming soon | Coming soon |
| ARC-AGI-1 | Coming soon | 94.50% | Coming soon |
| ARC-AGI-2 | Coming soon | 60.4% | Coming soon |

### Multimodal & Grounded

- Winner: Directional only
- Kimi K2.5 public-lane score: 66.8 (#24/49)
- Kimi K3 public-lane score: 89.4 (#1/49)

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| MMMU-Pro | 78.5% | 81.6% | Kimi K3 |
| Video-MME | 87.4% | Coming soon | Coming soon |
| MMVU | 80.4% | Coming soon | Coming soon |
| VideoMMMU | 86.6% | Coming soon | Coming soon |
| OfficeQA Pro | Coming soon | 63.3% | Coming soon |
| MMMU-Pro w/ Python | Coming soon | 83.4% | Coming soon |
| CharXiv w/o tools | Coming soon | 84.8% | Coming soon |
| CharXiv | Coming soon | 91.3% | Coming soon |
| MathVision | Coming soon | 94.3% | Coming soon |
| MathVision w/ Python | Coming soon | 97.8% | Coming soon |
| BabyVision w/ Python | Coming soon | 85.7% | Coming soon |
| ZeroBench | Coming soon | 23.0% | Coming soon |
| ZeroBench w/ Python | Coming soon | 41.0% | Coming soon |
| WorldVQA ForceAnswer | Coming soon | 51.0% | Coming soon |
| OmniDocBench | Coming soon | 91.1% | Coming soon |
| PerceptionBench | Coming soon | 58.5% | Coming soon |

### Knowledge

- Winner: Directional only
- Kimi K2.5 public-lane score: 49.5 (Estimated · #66/171)
- Kimi K3 public-lane score: 67.9 (Supported · #18/171)

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| GPQA | 87.6% | 93.5% | Kimi K3 |
| GPQA-D | 87.6% | 93.5% | Kimi K3 |
| SuperGPQA | 69.2% | Coming soon | Coming soon |
| MMLU-Pro | 87.1% | Coming soon | Coming soon |
| MMLU-Pro (Arcee) | 87.1% | Coming soon | Coming soon |
| HLE | 30.1% | 56% | Kimi K3 |
| HLE w/o tools | Coming soon | 43.5% | Coming soon |
| GPQA Diamond (Vals) | Coming soon | 92.9% | Coming soon |
| MMLU-Pro (Vals) | Coming soon | 88.0% | Coming soon |

### Multilingual

- Winner: Not comparable
- Kimi K2.5 public-lane score: 87.3 (#8/16)
- Kimi K3 public-lane score: Coming soon

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| MMLU-ProX | 82.3% | Coming soon | Coming soon |
| NOVA-63 | 56.0% | Coming soon | Coming soon |

### Instruction Following

- Winner: Not comparable
- Kimi K2.5 public-lane score: 84.4 (#41/125)
- Kimi K3 public-lane score: Coming soon

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| IFEval | 93.9% | Coming soon | Coming soon |
| Gray Swan IPI (15 attempts) | Coming soon | 52.7% | Coming soon |

### Mathematics

- Winner: Not comparable
- Kimi K2.5 public-lane score: 62.1 (#4/7)
- Kimi K3 public-lane score: Coming soon

| Benchmark | Kimi K2.5 | Kimi K3 | Winner |
|-----------|-----------|-----------|--------|
| AIME 2025 | 96.1% | Coming soon | Coming soon |
| AIME26 | 95.8% | Coming soon | Coming soon |
| AIME25 (Arcee) | 96.3% | Coming soon | Coming soon |
| HMMT Feb 2025 | 95.4% | Coming soon | Coming soon |
| HMMT Nov 2025 | 91.1% | Coming soon | Coming soon |
| HMMT Feb 2026 | 87.1% | Coming soon | Coming soon |
| MMAnswerBench | 81.8% | Coming soon | Coming soon |
| FrontierMath v2 (Tiers 1-3) | 27.900% | Coming soon | Coming soon |
| FrontierMath v2 (Tier 4) | 4.200% | Coming soon | Coming soon |

## FAQ

### Which is better overall, Kimi K2.5 or Kimi K3?

Kimi K3 is ahead overall on BenchLM right now.

### Where is the biggest gap between Kimi K2.5 and Kimi K3?

The widest category gap is in agentic, where the averages are 38 for Kimi K2.5 and 68.1 for Kimi K3.

### How many benchmarks does Kimi K2.5 cover on BenchLM?

Kimi K2.5 currently has 45 sourced benchmark scores on BenchLM.

### How many benchmarks does Kimi K3 cover on BenchLM?

Kimi K3 currently has 48 sourced benchmark scores on BenchLM.

## Related Comparisons

- [Kimi K2.5 vs Kimi K2.5 (Reasoning)](/compare/kimi-k2-5-vs-kimi-k2-5-reasoning)
- [Kimi K3 vs Kimi K2.5 (Reasoning)](/compare/kimi-k2-5-reasoning-vs-kimi-k3)
- [Kimi K2.5 vs GPT-6 Astra](/compare/gpt-6-astra-vs-kimi-k2-5)
- [Kimi K3 vs GPT-6 Astra](/compare/gpt-6-astra-vs-kimi-k3)

## Explore More

- [Kimi K2.5 profile](/models/kimi-k2-5)
- [Kimi K3 profile](/models/kimi-k3)
- [Compare Pricing](/llm-pricing)
- [Alternative Finder](/tools/alternative-finder)
- [LLM Selector](/tools/llm-selector)
- [Overall Rankings](/best/overall)