Model

Gemma 2 9B

OPEN

Google · Gemma 2 family · released Jun 27, 2024

Efficient mid-size Gemma 2 with GQA, online-distilled; outperforms Llama 3 8B in its size class.

ReasoningCodingVisionFunction callingTool useAgentic

479.2

Elo · rank #311

Parameters

9.24B

Active params

9.24B (dense)

Context

8K tokens

Architecture

Dense decoder-only Transformer, GQA + interleaved local/global attention, distilled from a larger model

License

Gemma Terms of Use

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AGIEvalReasoning52.8%#14

best: OLMo 3-Think 32B · 88.2%

ARC-ChallengeReasoning68.4%#55

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning88.0%#10

best: Phi-3-medium (14B) · 97.7%

Arena EloHuman preference1187#43

best: Claude Fable 5 · 1505

BIG-Bench HardReasoning68.2%#64

best: ERNIE 4.5 300B-A47B · 94.3%

DROPReasoning69.4#45

best: Hunyuan-T1 · 93.1

GSM8KMath68.6%#102

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning81.9%#68

best: Claude 3 Opus · 95.4%

HumanEvalCoding40.2%#137

best: Claude Opus 4.5 · 99.4%

KoMT-BenchHuman preference7.92#3

best: LG EXAONE 3.0 7.8B Instruct · 8.92

MATH-500Math36.6%#150

best: GPT-5 · 99.4%

MBPPCoding52.4%#70

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MMLUKnowledge71.3%#138

best: OpenAI o3 · 92.9%

PIQAReasoning81.7%#31

best: GPT-4o mini · 93.1%

Social IQaReasoning53.4%#16

best: Apple DCLM-Baseline 7B · 82.9%

TriviaQAKnowledge76.6%#16

best: Sarvam-1 (2B) · 90.6%

WinoGrandeReasoning80.6%#34

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

7 GB

VRAM @ FP16

20 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF Q4 · GGUF Q8 · AWQ · MLX

Gemma 2 family

Elo progression across releases

Gemma 2 27BJun 2024628.2 Gemma 2 9BJun 2024479.2

Resources

HFgoogle/gemma-2-9b GHgoogle-deepmind/gemma DOCSai.google.dev/gemma/docs/core/model_card_2

API price weights · each benchmark row carries its own source badge (see methodology)