Inference
EN
Google

Gemma 4 12B

Gemma 4 12B is a reasoning model from Google in the Gemma 4 family. 16 benchmarks count toward its index, in 8 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2026-06
RankedEnough results for an index.
IndexAll categories, out of 100.39.0±4.0Median 51.0
CoverageIndex weight that has results.100%
SpeedOutput tokens per second.—Median 67/s
First tokenSeconds to the first output token.—Median 1.3 s
Input / 1MUS dollars per 1M input tokens.—Median $0.60
Output / 1MUS dollars per 1M output tokens.—Median $2.55
ContextMaximum tokens in one request.256KMedian 262K
EloLMArena rating and rank.—Median 1422

Run locally

11.96B parameters
16-bit
26.1 GB
8-bit
14.3 GB
4-bit
8.6 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy fits in 12 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.2 counted30.5200 / 218
CodingCode writing and repair.1 counted43.5—
ReasoningLogic problems and puzzles.3 counted41.5154 / 227
MultimodalTasks with images and text.3 counted36.277 / 84
KnowledgeFacts and expert knowledge.4 counted41.5183 / 268
MultilingualTasks in many languages.1 counted29.7—
InstructionTasks with strict prompt rules.1 counted57.7—
MathMath problems.1 counted41.1—

Results

16 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
Artificial Analysis IFBenchInstructionAbove fitFar above this model's index.73.5%28 / 146
EuroEval PortugueseMultilingual45.5% ±0.983 / 182
EuroEval DutchMultilingual44.9% ±0.784 / 175
EuroEval PolishMultilingual43.1% ±0.8101 / 209
EuroEval GermanMultilingual37.3% ±1.093 / 191
EuroEval SwedishMultilingual42.0% ±1.097 / 195
Graduate-Level Google-Proof Q&AKnowledge78.8%71 / 141
EuroEval ItalianMultilingual41.5% ±0.992 / 174
Artificial Analysis Humanity's Last ExamKnowledge15.7%126 / 222
Artificial Analysis GPQA DiamondKnowledge75.3%115 / 199
EuroEval SpanishMultilingual32.3% ±0.9104 / 176
MMMLUKnowledge83.4%9 / 15
Massive Multitask Language Understanding ProfessionalKnowledge77.2%45 / 75
EuroEval FrenchMultilingual42.8% ±0.9108 / 178
Artificial Analysis Long Context ReasoningReasoning63.7%137 / 217
LiveCodeBench v6Coding72.0%39 / 59
Artificial Analysis MMMU-ProMultimodal69.7%76 / 113
Critical Physics TasksReasoning0.0%150 / 214
τ²-Bench Tool-Agent-User EvaluationAgentic36.3%108 / 149
GDPval-AA normalizedAgentic0.0%101 / 133
GPQA Diamond · BenchLMKnowledge78.8%64 / 84
MathVisionMultimodal79.7%20 / 24
Massive Multi-discipline Multimodal Understanding ProMultimodal69.1%45 / 53
AIME 2026Math77.5%29 / 34
MRCRv2Reasoning43.4%12 / 14
ExtractBench: schema-guided enterprise document extraction (LlamaIndex leaderboard)Multimodal45.4%26 / 26
Humanity's Last Exam without toolsKnowledge5.2%53 / 53
MedXpertQA MultimodalMultimodal48.7%—

Sources

BenchLMCC BY-NC 4.0EuroEvalMIT

History

5 releases
Index39.2 → 39.0
23 Oct 202612 Nov 2026

Questions

How good is Gemma 4 12B?
Its index is 39.0 out of 100. That is place 214 of 283 ranked models.
What is Gemma 4 12B best at?
Reasoning: 41.5 out of 100, place 154 of 227.
What is the context window of Gemma 4 12B?
256K tokens.
Is Gemma 4 12B open source?
Yes. Its weights are published.
When was Gemma 4 12B released?
June 2026.

Badge

Gemma 4 12B on Inference360: place 214 of 283, index 39.0
HTML
<a href="https://inference360.ai/models/gemma-4-12b"><img src="https://inference360.ai/badges/gemma-4-12b.svg" alt="Gemma 4 12B on Inference360: place 214 of 283, index 39.0" height="20"></a>
Markdown
[![Gemma 4 12B on Inference360: place 214 of 283, index 39.0](https://inference360.ai/badges/gemma-4-12b.svg)](https://inference360.ai/models/gemma-4-12b)

Compare with

Solar Pro 339.2DeepSeek-R138.9Gemini 3.6 Flash63.7Gemini 3.5 Flash64.2

More from Google

All models
Gemini 3.5 Flash64.2Nano Banana Pro (Gemini 3 Pro Image)UnrankedNano Banana 2 (Gemini 3.1 Flash Image)UnrankedNano Banana 2 Lite (Gemini 3.1 Flash Lite Image)UnrankedGemini Pro LatestUnranked