Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
Gemma 4 26B A4B
64
1/8 categoriesKimi K2.5
72
Winner · 3/8 categoriesGemma 4 26B A4B· Kimi K2.5
Pick Kimi K2.5 if you want the stronger benchmark profile. Gemma 4 26B A4B only becomes the better choice if coding is the priority or you want the cheaper token bill.
Kimi K2.5 is clearly ahead on the aggregate, 72 to 64. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
Kimi K2.5's sharpest advantage is in reasoning, where it averages 66.9 against 44.1. The single biggest benchmark swing on the page is MRCRv2, 44.1% to 70%. Gemma 4 26B A4B does hit back in coding, so the answer changes if that is the part of the workload you care about most.
Kimi K2.5 is also the more expensive model on tokens at $0.50 input / $2.80 output per 1M tokens, versus $0.00 input / $0.00 output per 1M tokens for Gemma 4 26B A4B. That is roughly Infinityx on output cost alone. Gemma 4 26B A4B is the reasoning model in the pair, while Kimi K2.5 is not. That usually helps on harder chain-of-thought-heavy tests, but it can also mean more latency and more token spend in real use. Gemma 4 26B A4B gives you the larger context window at 256K, compared with 128K for Kimi K2.5.
BenchLM keeps the benchmark table and the operator tradeoffs on the same page so a better score does not hide a materially slower, pricier, or smaller-context model.
Runtime metrics show N/A when BenchLM does not have a sourced snapshot for that exact model. The scoring rules and freshness policy are documented on the methodology page.
| Benchmark | Gemma 4 26B A4B | Kimi K2.5 |
|---|---|---|
| Agentic | ||
| Terminal-Bench 2.0 | — | 50.8% |
| BrowseComp | — | 60.6% |
| OSWorld-Verified | — | 63.3% |
| Claw-Eval | — | 52.9% |
| QwenClawBench | — | 54.3% |
| QwenWebBench | — | 1160 |
| TAU3-Bench | — | 65.7% |
| VITA-Bench | — | 36.0% |
| DeepPlanning | — | 14.4% |
| Toolathlon | — | 27.8% |
| MCP Atlas | — | 29.5% |
| MCP-Tasks | — | 59.1% |
| WideResearch | — | 72.7% |
| BrowseComp-VL | — | 42.9% |
| OSWorld | — | 63.3% |
| Tau2-Airline | — | 80.0% |
| Tau2-Telecom | — | 95.9% |
| PinchBench | — | 84.8% |
| BFCL v4 | — | 68.3% |
| AndroidWorld | — | 43.1% |
| WebVoyager | — | 84.3% |
| CodingGemma 4 26B A4B wins | ||
| LiveCodeBench | 77.1% | 85% |
| HumanEval | — | 99% |
| SWE-bench Verified | — | 76.8% |
| SWE-bench Verified* | — | 70.8% |
| LiveCodeBench v6 | — | 85.6% |
| SWE-bench Pro | — | 53.8% |
| SWE Multilingual | — | 73% |
| SWE-Rebench | — | 58.5% |
| React Native Evals | — | 74.9% |
| NL2Repo | — | 32% |
| Multimodal & GroundedKimi K2.5 wins | ||
| MMMU-Pro | 73.8% | 78.5% |
| MMMU | — | 84.3% |
| OfficeQA Pro | — | 69% |
| RealWorldQA | — | 81.0% |
| OmniDocBench 1.5 | — | 88.8% |
| Video-MME (with subtitle) | — | 87.4% |
| Video-MME (w/o subtitle) | — | 83.2% |
| MathVision | — | 84.2% |
| We-Math | — | 84.7% |
| DynaMath | — | 84.4% |
| MStar | — | 80.5% |
| Design2Code | — | 91.3% |
| Flame-VLM-Code | — | 88.8% |
| Vision2Web | — | 33.2% |
| ImageMining | — | 24.4% |
| MMSearch | — | 58.7% |
| MMSearch-Plus | — | 25.6% |
| SimpleVQA | — | 71.2% |
| ChatCVQA | — | 77.5% |
| MMLongBench-Doc | — | 58.5% |
| CC-OCR | — | 79.7% |
| AI2D_TEST | — | 90.8% |
| CountBench | — | 94.1% |
| RefCOCO (avg) | — | 87.8% |
| VideoMMMU | — | 86.6% |
| MLVU (M-Avg) | — | 85.0% |
| Facts-VLM | — | 57.8% |
| V* | — | 84.3% |
| ReasoningKimi K2.5 wins | ||
| BBH | 64.8% | 81% |
| MRCRv2 | 44.1% | 70% |
| MuSR | — | 72% |
| LongBench v2 | — | 61% |
| AI-Needle | — | 70% |
| KnowledgeKimi K2.5 wins | ||
| GPQA | 82.3% | 87.6% |
| MMLU-Pro | 82.6% | 87.1% |
| HLE | 17.2% | 30.1% |
| HLE w/o tools | 8.7% | — |
| MMLU | — | 77% |
| GPQA-D | — | 86.9% |
| SuperGPQA | — | 69.2% |
| MMLU-Pro (Arcee) | — | 87.1% |
| MMLU-Redux | — | 94.5% |
| C-Eval | — | 94% |
| FrontierScience | — | 67% |
| SimpleQA | — | 74% |
| Instruction Following | ||
| IFEval | — | 93.9% |
| IFBench | — | 70.2% |
| Multilingual | ||
| MGSM | — | 83% |
| MMLU-ProX | — | 82.3% |
| NOVA-63 | — | 56.0% |
| INCLUDE | — | 83.3% |
| PolyMath | — | 43.1% |
| VWT2k-lite | — | 77.6% |
| MAXIFE | — | 72.8% |
| Mathematics | ||
| AIME 2023 | — | 77% |
| AIME 2024 | — | 79% |
| AIME 2025 | — | 78% |
| AIME26 | — | 95.8% |
| AIME25 (Arcee) | — | 96.3% |
| HMMT Feb 2023 | — | 73% |
| HMMT Feb 2024 | — | 75% |
| HMMT Feb 2025 | — | 74% |
| HMMT Feb 2025 | — | 95.4% |
| HMMT Nov 2025 | — | 91.1% |
| HMMT Feb 2026 | — | 87.1% |
| MMAnswerBench | — | 81.8% |
| BRUMO 2025 | — | 76% |
| MATH-500 | — | 82% |
Kimi K2.5 is ahead overall, 72 to 64. The biggest single separator in this matchup is MRCRv2, where the scores are 44.1% and 70%.
Kimi K2.5 has the edge for knowledge tasks in this comparison, averaging 66.6 versus 56.1. Inside this category, HLE is the benchmark that creates the most daylight between them.
Gemma 4 26B A4B has the edge for coding in this comparison, averaging 77.1 versus 66.7. Inside this category, LiveCodeBench is the benchmark that creates the most daylight between them.
Kimi K2.5 has the edge for reasoning in this comparison, averaging 66.9 versus 44.1. Inside this category, MRCRv2 is the benchmark that creates the most daylight between them.
Kimi K2.5 has the edge for multimodal and grounded tasks in this comparison, averaging 74.2 versus 73.8. Inside this category, MMMU-Pro is the benchmark that creates the most daylight between them.
Get notified when new models drop, benchmark scores change, or the leaderboard shifts. One email per week.
Free. No spam. Unsubscribe anytime. We only store derived location metadata for consent routing.