Gemma 4 31B vs Kimi K2.5
Wins 1 of 8 areas
Following instructions
Wins 7 of 8 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math · Long documents
Kimi K2.5 is the stronger all-rounder.Gemma 4 31B is cheaper and better at following instructions.
Scores updated · 49 tests both models report · How we compare
Where each one wins
Tests won in each of the eight areas we test. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software17Kimi K2.57 of 8 tests
- Images and chartsUnderstanding pictures, charts and video03Kimi K2.53 of 4 tests · 1 tie
- ReasoningHard problems that need careful thinking03Kimi K2.53 of 3 tests
- MathCompetition and research-level math03Kimi K2.53 of 3 tests
- FactsGetting facts right instead of making them up13Kimi K2.53 of 4 tests
- AgentsCarrying out multi-step tasks on its own02Kimi K2.52 of 3 tests · 1 tie
- Long documentsFinding answers in very long texts01Kimi K2.51 of 1 test
- Following instructionsDoing exactly what it is asked10Gemma 4 31B1 of 1 test
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
Gemma 4 31B costs 84% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where Kimi K2.5 pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+24.8points ahead
- Short factual questions, answered correctlySimpleQA Verified+23.9points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+21.3points ahead
Where Gemma 4 31B pulls ahead
- Follows unfamiliar, precisely checkable instructionsIFBench+5.4points ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+1.6points ahead
Every test, side by side
All 49 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingKimi K2.5
- SWE-bench VerifiedKimi K2.5 by 24.85276.8+24.8
- SWE-bench MultilingualKimi K2.5 by 21.351.773+21.3
- SWE-bench ProKimi K2.5 by 1535.750.7+15
- LMArena · WebDevKimi K2.5 by 71 rating points13651436+71 rating
- LiveCodeBench v6Kimi K2.5 by 58085+5
- SciCodeKimi K2.5 by 3.545.549+3.5
- Terminal-Bench 2.1Kimi K2.5 by 2.343.445.7+2.3
- Terminal-Bench HardGemma 4 31B by 1.636.434.8+1.6
AgentsKimi K2.5
- GDPValKimi K2.5 by 11.16.117.2+11.1
- MCP AtlasKimi K2.5 by 6.657.263.8+6.6
- τ-Bench V3 · Bankingtie14.814.2tie
ReasoningKimi K2.5
- Humanity's Last ExamKimi K2.5 by 7.123.630.7+7.1
- GPQA DiamondKimi K2.5 by 2.285.787.9+2.2
- CritPtKimi K2.5 by 1.71.43.1+1.7
FactsKimi K2.5
- SimpleQA VerifiedKimi K2.5 by 23.910.434.3+23.9
- AA-Omniscience · Non-hallucinationKimi K2.5 by 19.31534.3+19.3
- AA-Omniscience · AccuracyKimi K2.5 by 15.22035.2+15.2
- Vectara HHEM hallucination ratelower is betterGemma 4 31B by 6.87.414.2+6.8
Images and chartsKimi K2.5
- MathVistaKimi K2.5 by 10.879.390.1+10.8
- CharXiv (RQ)Kimi K2.5 by 9.667.977.5+9.6
- MMMU-ProKimi K2.5 by 273.475.4+2
- LMArena · Visiontie12771269tie
MathKimi K2.5
- HMMT Feb. 2026Kimi K2.5 by 9.977.287.1+9.9
- IMOAnswerBenchKimi K2.5 by 7.374.581.8+7.3
- AIME 2026Kimi K2.5 by 6.689.295.8+6.6
Following instructionsGemma 4 31B
- IFBenchGemma 4 31B by 5.475.670.2+5.4
Other results22 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- WideSearchKimi K2.5 by 43.835.279+43.8
- AA-OmniscienceKimi K2.5 by 40.6-47.9-7.3+40.6
- τ²-Bench Telecom (AA run)Kimi K2.5 by 3659.995.9+36
- HLE (with tools)Kimi K2.5 by 23.726.550.2+23.7
- SimpleVQAKimi K2.5 by 18.352.971.2+18.3
- NL2RepoKimi K2.5 by 16.515.532+16.5
- AA Agentic IndexKimi K2.5 by 156.721.7+15
- MCPMarkKimi K2.5 by 11.418.129.5+11.4
- AA IntelligenceKimi K2.5 by 8.814.723.5+8.8
- OmniDocBench 1.5Kimi K2.5 by 8.780.188.8+8.7
- τ²-BenchKimi K2.5 by 8.576.985.4+8.5
- Terminal-Bench 2.0Kimi K2.5 by 7.942.950.8+7.9
- vectara_answer_rateGemma 4 31B by 7.810092.2+7.8
- vectara_factual_consistencyGemma 4 31B by 6.892.685.8+6.8
- HMMT Feb. 2025Kimi K2.5 by 6.788.795.4+6.7
- Tool-DecathlonKimi K2.5 by 6.621.227.8+6.6
- vectara_avg_summary_lengthKimi K2.5 by 36.2 rating points75.8112+36.2 rating
- Artificial Analysis Coding IndexKimi K2.5 by 3.443.446.8+3.4
- MMLU-ProKimi K2.5 by 1.985.287.1+1.9
- HMMT Nov. 2025Kimi K2.5 by 1.787.589.2+1.7
- τ³-BenchGemma 4 31B by 1.567.566+1.5
- MATH-VisionGemma 4 31B by 1.485.684.2+1.4
Questions people ask
Which is better, Gemma 4 31B or Kimi K2.5?
Kimi K2.5 wins seven of the eight areas we test: coding, agents, reasoning, facts, images and charts, math and long documents. Gemma 4 31B wins following instructions, and costs 84% less.
Which is better for coding?
Kimi K2.5. It wins 7 of the 8 coding tests both models report; Gemma 4 31B wins 1.
Which is cheaper?
Gemma 4 31B costs $0.17 per million input tokens and $0.40 per million output tokens; Kimi K2.5 costs $0.60 and $3.00. That makes Gemma 4 31B about 84% cheaper for the same work.
How do you compare the two?
We use the 49 benchmark tests both models have published scores on. The verdict counts the 27 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 22 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.