Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
Gemma 4 31B
73
Winner · 2/8 categoriesKimi K2.5
72
2/8 categoriesGemma 4 31B· Kimi K2.5
Pick Gemma 4 31B if you want the stronger benchmark profile. Kimi K2.5 only becomes the better choice if knowledge is the priority or you would rather avoid the extra latency and token burn of a reasoning model.
Gemma 4 31B finishes one point ahead overall, 73 to 72. That is enough to call, but not enough to treat as a blowout. This matchup comes down to a few meaningful edges rather than one model dominating the board.
Gemma 4 31B's sharpest advantage is in coding, where it averages 80 against 66.7. The single biggest benchmark swing on the page is BBH, 74.4% to 81%. Kimi K2.5 does hit back in knowledge, 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 31B. That is roughly Infinityx on output cost alone. Gemma 4 31B 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 31B 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 31B | 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 31B wins | ||
| LiveCodeBench | 80% | 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 & GroundedGemma 4 31B wins | ||
| MMMU-Pro | 76.9% | 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 | 74.4% | 81% |
| MRCRv2 | 66.4% | 70% |
| MuSR | — | 72% |
| LongBench v2 | — | 61% |
| AI-Needle | — | 70% |
| KnowledgeKimi K2.5 wins | ||
| GPQA | 84.3% | 87.6% |
| MMLU-Pro | 85.2% | 87.1% |
| HLE | 26.5% | 30.1% |
| HLE w/o tools | 19.5% | — |
| 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% |
Gemma 4 31B is ahead overall, 73 to 72. The biggest single separator in this matchup is BBH, where the scores are 74.4% and 81%.
Kimi K2.5 has the edge for knowledge tasks in this comparison, averaging 66.6 versus 61.3. Inside this category, HLE is the benchmark that creates the most daylight between them.
Gemma 4 31B has the edge for coding in this comparison, averaging 80 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 66.4. Inside this category, BBH is the benchmark that creates the most daylight between them.
Gemma 4 31B has the edge for multimodal and grounded tasks in this comparison, averaging 76.9 versus 74.2. 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.