← Back to Directory
Google DeepMind
LLM📅 Released: 2025-05-10

Gemma 3

Gemma 3 is a highly efficient open model from Google.

#open-source#efficient#on-device

Overview

Gemma 3 is Google's 2025 contribution to the open-source community, built from the same research as the Gemini family. It is a multimodal-native small model that excels in on-device reasoning, making it the perfect foundation for privacy-first, local AI applications.

Unique Factor

The world's most capable open model under 30B parameters with native vision support.

Key Capabilities

On-device optimized
Native vision
Open weight

Benchmarks

MMLU Score
83%
HumanEval (Coding)
82%
GPQA Diamond
65%
MATH Benchmark
78%

Top Use Cases

Local Privacy Assistant

Helping users manage local data (emails, files) without cloud access.

Example: “Scan my recent downloads folder and summarize the PDF files.

Detailed Features

01

Multimodal Native: Can see and understand images entirely on local hardware.

02

High Reasoning Efficiency: Outperforms models twice its size on math and logic benchmarks.

03

On-device Optimized: Designed to run efficiently on high-end smartphones and AI PCs.

04

Open Weights (Gemma License): Fully accessible for fine-tuning and commercial use.

05

Google Ecosystem Ready: Native support for Vertex AI, Keras, and PyTorch.

06

Strong Multilingual Base: Trained on a diverse global dataset for cross-cultural utility.

Strengths & Pros

  • Incredible performance for its size
  • Native multimodal capability
  • Free and open to use

Limitations & Cons

  • Limited context window (32k)
  • Smaller knowledge base than 70B+ models

Ideal Usage & Target Audience

Best For

Mobile developers, local AI researchers, and edge compute engineers.

Not Recommended For

Users needing deep scientific research or massive context.

API Implementation

python
# Running Gemma 3 locally
from transformers import pipeline
pipe = pipeline('text-generation', model='google/gemma-3-27b')
print(pipe('What is quantum entanglement?'))

Check the official documentation for full SDK details.

Learn to Master This Model

Take our free structured Gemma course — from basics to advanced techniques.

Gemini Course

Technical Specs

Context32,768 tokens
Params27B
LicenseGemma Terms
ArchTransformer

API Pricing

$0 / 1M input tokens

Output: $0 / 1M tokens

✓ Free tier available
Access API

Developer

The scientific leaders of AI — creators of Gemini and the innovators behind the Transformer architecture.

Prompt Library

Browse Coding Prompts

📋

Previous Version

Gemma 2