Model

Gemma 7B

OPEN

Google · Gemma family · released Feb 21, 2024

First open-weights model built from Gemini research; 8.54B params, 8K context, strong for its size on MMLU/HellaSwag.

ReasoningCodingVisionFunction callingTool useAgentic

186.8

Elo · rank #365

Parameters

8.54B

Active params

8.54B (dense)

Context

8K tokens

Architecture

Dense decoder-only transformer (28 layers, d_model 3072, 256k vocab)

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

AGIEvalReasoning41.7%#30

best: OLMo 3-Think 32B · 88.2%

ARC-ChallengeReasoning53.2%#112

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning81.5%#23

best: Phi-3-medium (14B) · 97.7%

BIG-Bench HardReasoning55.1%#85

best: ERNIE 4.5 300B-A47B · 94.3%

GSM8KMath46.4%#136

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning81.2%#71

best: Claude 3 Opus · 95.4%

HumanEvalCoding32.3%#148

best: Claude Opus 4.5 · 99.4%

MATH-500Math24.3%#173

best: GPT-5 · 99.4%

MBPPCoding44.4%#79

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

MMLUKnowledge64.3%#181

best: OpenAI o3 · 92.9%

OpenBookQAReasoning52.8%#22

best: Claude 1 · 90.8%

PIQAReasoning81.2%#33

best: GPT-4o mini · 93.1%

Social IQaReasoning51.8%#21

best: Apple DCLM-Baseline 7B · 82.9%

TriviaQAKnowledge63.4%#32

best: Sarvam-1 (2B) · 90.6%

TruthfulQAKnowledge31.8%#102

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

WinoGrandeReasoning72.3%#86

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

6 GB

VRAM @ FP16

18 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_K_M · GGUF Q8_0 · AWQ · MLX

Gemma family

Elo progression across releases

Gemma 2BFeb 2024-150.7 Gemma 7BFeb 2024186.8

Resources

HFgoogle/gemma-7b GHgoogle-deepmind/gemma DOCSai.google.dev/gemma/docs/releases APIAvailable via Google API

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