Inference
EN
Google

Gemma 4 31B

Gemma 4 31B is a reasoning model from Google in the Gemma 4 family. 24 benchmarks count toward its index, in 8 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2026-04
RankedEnough results for an index.
IndexAll categories, out of 100.47.9±3.3Median 51.0
CoverageIndex weight that has results.100%
SpeedOutput tokens per second.35/sMedian 67/s
First tokenSeconds to the first output token.50.0 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.FreeMedian $0.60
Output / 1MUS dollars per 1M output tokens.FreeMedian $2.55
ContextMaximum tokens in one request.256KMedian 262K
EloLMArena rating and rank.1443±7 #73Median 1422

Run locally

31.27B parameters
16-bit
66.7 GB
8-bit
35.9 GB
4-bit
20.9 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy fits in 24 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.6 counted44.9162 / 218
CodingCode writing and repair.4 counted50.1136 / 209
ReasoningLogic problems and puzzles.4 counted46.6139 / 227
MultimodalTasks with images and text.2 counted56.949 / 84
KnowledgeFacts and expert knowledge.5 counted44.6169 / 268
MultilingualTasks in many languages.1 counted47.5—
InstructionTasks with strict prompt rules.1 counted59.7—
MathMath problems.1 counted51.6—

Results

24 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
EuroEval PortugueseMultilingual67.5% ±0.612 / 182
EuroEval ItalianMultilingual66.8% ±0.412 / 174
EuroEval FrenchMultilingual67.9% ±0.414 / 178
EuroEval PolishMultilingual61.6% ±0.417 / 209
EuroEval GermanMultilingual55.7% ±0.718 / 191
EuroEval DutchMultilingual62.7% ±0.418 / 175
EuroEval SwedishMultilingual63.1% ±0.826 / 195
Artificial Analysis IFBenchInstruction75.6%21 / 146
EuroEval SpanishMultilingual55.8% ±0.726 / 176
Massive Multitask Language Understanding ProfessionalKnowledge85.2%18 / 75
EuroEval PortugueseMultilingual55.3% ±0.751 / 182
EuroEval SwedishMultilingual50.3% ±1.266 / 195
EuroEval ItalianMultilingual50.3% ±0.961 / 174
EuroEval DutchMultilingual51.8% ±0.762 / 175
Artificial Analysis GPQA DiamondKnowledge85.7%72 / 199
EuroEval GermanMultilingual42.6% ±0.870 / 191
EuroEval PolishMultilingual46.7% ±0.978 / 209
Graduate-Level Google-Proof Q&AKnowledge84.3%54 / 141
EuroEval SpanishMultilingual40.8% ±0.871 / 176
Artificial Analysis Humanity's Last ExamKnowledge23.6%103 / 222
EuroEval FrenchMultilingual50.0% ±0.883 / 178
OTIS Mock AIME 2024-2025Math73.3% ±6.794 / 200
GPQA diamond · Epoch AIKnowledge75.8% ±3.1103 / 217
Artificial Analysis Long Context ReasoningReasoning69.7%113 / 217
LiveCodeBench v6Coding80.0%31 / 59
Critical Physics TasksReasoning1.4%113 / 214
Terminal-Bench 2.0 · Vals AIAgentic39.3% ±5.235 / 66
Massive Multi-discipline Multimodal Understanding ProMultimodal76.9%30 / 53
Artificial Analysis MMMU-ProMultimodal73.4%64 / 113
τ²-Bench Tool-Agent-User EvaluationAgentic59.9%87 / 149
Artificial Analysis SciCodeCoding45.5%70 / 114
ParseBench: document parsing for AI agents (LlamaIndex leaderboard)Multimodal62.4%24 / 39
React Native EvalsCoding75.2%10 / 16
MRCRv2Reasoning66.4%9 / 14
GDPval-AA normalizedAgentic5.9%95 / 133
Humanity's Last ExamKnowledge26.5%56 / 77
Artificial Analysis Omniscience AccuracyKnowledge19.9%131 / 164
Gert Labs Composite Game BenchmarkAgentic35.3%45 / 52
LiveBench Long AgenticAgentic13.3%46 / 51
Artificial Analysis ITBench-AAAgentic37.3%22 / 24
LiveBench ResearchAgentic44.4%48 / 52
Humanity's Last Exam without toolsKnowledge19.5%49 / 53
LiveBench DesignAgentic58.4%49 / 52
LiveBench Coding 2.0Coding49.0%49 / 52
LiveBench Browser UseAgentic4.0%47 / 49
LiveBench ReasoningReasoning68.3%77 / 80
LiveBench Data AnalysisReasoning53.3%77 / 80
SimpleQA VerifiedKnowledge10.4% ±1.089 / 91
Finance Agent v1.1Finance50.8% ±2.927 / 51
MortgageTaxFinance61.4% ±1.060 / 97
CaseLaw v2Legal52.6% ±1.342 / 53
SAGEEducation82.8% ±2.124 / 29
Refusal of harmful cyber requests (JailbreakBench, StrongREJECT, AgentHarm)Agentic84.0%—
b3: Backbone Breaker BenchmarkAgentic83.0%—
SWE-RebenchCoding41.6%—

Sources

BenchLMCC BY-NC 4.0Epoch AICC BYLiveBenchNo licence statedEuroEvalMITVals AINo licence stated

History

5 releases
Index48.2 → 47.9
23 Oct 202612 Nov 2026
Output / 1MFree → Free
23 Oct 202612 Nov 2026

Questions

How much does Gemma 4 31B cost?
Input tokens are free, output tokens are free.
How good is Gemma 4 31B?
Its index is 47.9 out of 100. That is place 162 of 283 ranked models.
What is Gemma 4 31B best at?
Multimodal: 56.9 out of 100, place 49 of 84.
How fast is Gemma 4 31B?
35 output tokens per second, 50.0 s to the first token.
What is the context window of Gemma 4 31B?
256K tokens.
Is Gemma 4 31B open source?
Yes. Its weights are published.
When was Gemma 4 31B released?
April 2026.

Badge

Gemma 4 31B on Inference360: place 162 of 283, index 47.9
HTML
<a href="https://inference360.ai/models/gemma-4-31b"><img src="https://inference360.ai/badges/gemma-4-31b.svg" alt="Gemma 4 31B on Inference360: place 162 of 283, index 47.9" height="20"></a>
Markdown
[![Gemma 4 31B on Inference360: place 162 of 283, index 47.9](https://inference360.ai/badges/gemma-4-31b.svg)](https://inference360.ai/models/gemma-4-31b)

Compare with

Claude Sonnet 4.547.9GLM-5V-Turbo48.1Claude Opus 4.147.5Qwen3 235B 2507 (Reasoning)47.4Gemini 3.1 Flash Lite Preview49.6

More from Google

All models
Gemma 4 26B A4B44.5Gemma 4 E4B28.2Gemma 4 E2B21.9Lyria 3 Pro PreviewUnrankedLyria 3 Clip PreviewUnranked