Gemma 4 12B

Gemma 4 12B is a reasoning model from Google in the Gemma 4 family. 18 benchmarks count toward its score, in 8 categories.

availableShows if the model has enough results for an index.
IndexOverall score out of 100.41.7 ±3.8
CoverageShare of the index weight with results.100%
SpeedOutput tokens per second.
Input / 1MUS dollars per 1M input tokens.N/A
Output / 1MUS dollars per 1M output tokens.N/A
ContextMaximum tokens in one request.256K
EloLMArena rating and rank.N/A

The index is a score out of 100. The ± range shows how much it can change.

CapabilitiesScore per category, out of 100.

Out of 100
AgenticMulti-step tasks with tools.
30.2
CodingCode writing and repair.
41.7
ReasoningLogic problems and puzzles.
45.3
MultimodalTasks with images and text.
46.0
KnowledgeFacts and expert knowledge.
41.0
MultilingualTasks in many languages.
55.2
InstructionTasks with strict rules in the prompt.
63.7
MathMath problems.
41.1

Results

18 counted
BenchmarkThe test name.CategoryThe capability that the test measures.ResultThe score from the publisher.IndexThis result as a score out of 100.RunThe settings of the run.DateDate of the result.Published byThe source of the result.
MMMLUKnowledge83.4%OpenAI
MathVisionMultimodal79.7%Qwen
Graduate-Level Google-Proof Q&AKnowledge78.8%51.0David Rein et al.
GPQA DiamondKnowledge78.8%51.0David Rein et al.
AIME 2026Math77.5%41.1Qwen
Massive Multitask Language Understanding ProfessionalKnowledge77.2%42.1Yubo Wang et al.
Artificial Analysis GPQA DiamondKnowledge75.3%45.7Artificial Analysis
Artificial Analysis IFBenchInstruction73.5%63.7Artificial Analysis
LiveCodeBench v6Coding72.0%42.6LiveCodeBench maintainers
Artificial Analysis MMMU-ProMultimodal69.7%52.3Artificial Analysis
Massive Multi-discipline Multimodal Understanding ProMultimodal69.1%39.8MMMU-Pro authors
Artificial Analysis Long Context ReasoningReasoning63.7%52.3Artificial Analysis
BIG-Bench HardReasoning53.0%Mirac Suzgun et al.
MedXpertQA MultimodalMultimodal48.7%Meta AI
EuroEval PortugueseMultilingual45.5%59.1EuroEval
EuroEval DutchMultilingual44.9%58.3EuroEval
MRCRv2Reasoning43.4%OpenAI
EuroEval PolishMultilingual43.1%56.1EuroEval
EuroEval FrenchMultilingual42.8%55.7EuroEval
EuroEval SwedishMultilingual42.0%54.8EuroEval
EuroEval ItalianMultilingual41.5%54.1EuroEval
EuroEval GermanMultilingual37.3%48.9EuroEval
τ²-Bench Tool-Agent-User EvaluationAgentic36.3%24.9Victor Barres et al.
EuroEval SpanishMultilingual32.3%42.7EuroEval
Artificial Analysis Coding IndexCoding31.0%40.8Artificial Analysis
Artificial Analysis Humanity's Last ExamKnowledge15.7%41.6Artificial Analysis
Artificial Analysis Omniscience AccuracyKnowledge15.6%33.2Artificial Analysis
Artificial Analysis Intelligence IndexKnowledge14.2%40.1Artificial Analysis
Humanity's Last Exam without toolsKnowledge5.2%33.2OpenAI
GDPval-AA normalizedAgentic0.0%35.6Artificial Analysis
Critical Physics TasksReasoning0.0%38.4Artificial Analysis

18 benchmarks count, from 26 of 31 results. A grey row does not count. Too few models took that benchmark.

Sources

BenchLM benchmark aggregationUsed with attribution; per-benchmark results credited to their original authorsEuroEval, collected directlyMIT — the leaderboard site and its CSV routes are in the licensed repository

More from Google

Gemini 3.7 Flash72.3Gemini 3.8 Flash72.1Gemini 3.5 Flash67.3Gemini 3.6 Flash66.8Gemini 3.1 Pro64.7