Google DeepMind DeepMind
code
Build with Gemini
spark
Chat with Gemini

Gemma 3

A family of lightweight models with multimodal understanding and unparalleled multilingual capabilities for more intelligent applications.

Explore Gemma 3

keyboard_arrow_right
Read docs
keyboard_arrow_right
View tech report

Gemma 3 is the most capable model that can run on a single GPU or TPU. Efficient on workstations, laptops, and even smartphones, allowing developers to build responsible AI applications at scale.

Capabilities

article

Handle complex tasks

Gemma 3's 128K-token context window lets your applications process and understand vast amounts of information, enabling more sophisticated AI features.

translate

Multilingual communication

Unparalleled multilingual capabilities let you communicate effortlessly across countries and cultures. Develop applications that reach a global audience, with support for over 140 languages.

books_movies_and_music

Multimodal understanding

Easily build applications that analyze images, text, and video opening up new possibilities for interactive and intelligent applications.

Model variants

270M

Compact model designed for both task-specific fine-tuning and strong instruction-following.

1B

Lightweight text model, ideal for small applications.

4B

Balanced for performance and flexibility, with multimodal support.

12B

Strong language capabilities, designed for complex tasks.

27B

Enhanced understanding, great for sophisticated applications.

Performance

MMLU-Pro benchmark bar chart. Scores: Gemma 2 2b (15.6%), Gemma 2 9b (46.8%), Gemma 2 27b (56.9%), Gemma 3 1b (14.7%), Gemma 3 4b (43.6%), Gemma 3 12b (60.6%), Gemma 3 27b (67.5%). MMLU-Pro benchmark bar chart. Scores: Gemma 2 2b (15.6%), Gemma 2 9b (46.8%), Gemma 2 27b (56.9%), Gemma 3 1b (14.7%), Gemma 3 4b (43.6%), Gemma 3 12b (60.6%), Gemma 3 27b (67.5%).

MMLU-Pro

The MMLU benchmark is a test that measures the breadth of knowledge and problem-solving ability acquired by large language models during pretraining.

LiveCodeBench benchmark bar chart. Scores: Gemma 2 2b (1.2%), Gemma 2 9b (10.8%), Gemma 2 27b (20.4%), Gemma 3 1b (1.9%), Gemma 3 4b (12.6%), Gemma 3 12b (24.6%), Gemma 3 27b (29.7%). LiveCodeBench benchmark bar chart. Scores: Gemma 2 2b (1.2%), Gemma 2 9b (10.8%), Gemma 2 27b (20.4%), Gemma 3 1b (1.9%), Gemma 3 4b (12.6%), Gemma 3 12b (24.6%), Gemma 3 27b (29.7%).

LiveCodeBench

Assesses code generation capabilities on real-world coding problems from platforms like LeetCode and Codeforces.

Bird-SQL benchmark bar chart. Scores: Gemma 2 2b (12.2%), Gemma 2 9b (33.8%), Gemma 2 27b (46.7%), Gemma 3 1b (6.4%), Gemma 3 4b (36.3%), Gemma 3 12b (47.9%), Gemma 3 27b (54.4%). Bird-SQL benchmark bar chart. Scores: Gemma 2 2b (12.2%), Gemma 2 9b (33.8%), Gemma 2 27b (46.7%), Gemma 3 1b (6.4%), Gemma 3 4b (36.3%), Gemma 3 12b (47.9%), Gemma 3 27b (54.4%).

Bird-SQL

Tests a model's ability to translate natural language questions into complex SQL queries across various domains.

GPQA Diamond benchmark bar chart. Scores: Gemma 2 2b (24.7%), Gemma 2 9b (28.8%), Gemma 2 27b (34.3%), Gemma 3 1b (19.2%), Gemma 3 4b (30.8%), Gemma 3 12b (40.9%), Gemma 3 27b (42.4%). GPQA Diamond benchmark bar chart. Scores: Gemma 2 2b (24.7%), Gemma 2 9b (28.8%), Gemma 2 27b (34.3%), Gemma 3 1b (19.2%), Gemma 3 4b (30.8%), Gemma 3 12b (40.9%), Gemma 3 27b (42.4%).

GPQA Diamond

Challenges models with difficult questions written by Ph.D. holders in biology, physics, and chemistry.

SimpleQA benchmark bar chart. Scores: Gemma 2 2b (2.8%), Gemma 2 9b (5.3%), Gemma 2 27b (9.2%), Gemma 3 1b (2.2%), Gemma 3 4b (4%), Gemma 3 12b (6.3%), Gemma 3 27b (10%). SimpleQA benchmark bar chart. Scores: Gemma 2 2b (2.8%), Gemma 2 9b (5.3%), Gemma 2 27b (9.2%), Gemma 3 1b (2.2%), Gemma 3 4b (4%), Gemma 3 12b (6.3%), Gemma 3 27b (10%).

SimpleQA

Evaluates a model's ability to answer simple, factual questions with short phrases.

FACTS Grounding benchmark bar chart. Scores: Gemma 2 2b (43.8%), Gemma 2 9b (62%), Gemma 2 27b (62.4%), Gemma 3 1b (36.4%), Gemma 3 4b (70.1%), Gemma 3 12b (75.8%), Gemma 3 27b (74.9%). FACTS Grounding benchmark bar chart. Scores: Gemma 2 2b (43.8%), Gemma 2 9b (62%), Gemma 2 27b (62.4%), Gemma 3 1b (36.4%), Gemma 3 4b (70.1%), Gemma 3 12b (75.8%), Gemma 3 27b (74.9%).

FACTS Grounding

Evaluates if LLM responses are factually accurate and detailed enough, based on given input documents.

MATH benchmark bar chart. Scores: Gemma 2 2b (27.2%), Gemma 2 9b (49.4%), Gemma 2 27b (55.6%), Gemma 3 1b (48%), Gemma 3 4b (75.6%), Gemma 3 12b (83.8%), Gemma 3 27b (89%). MATH benchmark bar chart. Scores: Gemma 2 2b (27.2%), Gemma 2 9b (49.4%), Gemma 2 27b (55.6%), Gemma 3 1b (48%), Gemma 3 4b (75.6%), Gemma 3 12b (83.8%), Gemma 3 27b (89%).

MATH

MATH evaluates a language model's ability to solve complex mathematical word problems, requiring reasoning, multi-step problem-solving, and the understanding of mathematical concepts.

HiddenMath benchmark bar chart. Scores: Gemma 2 2b (1.8%), Gemma 2 9b (10.4%), Gemma 2 27b (14.8%), Gemma 3 1b (15.8%), Gemma 3 4b (43%), Gemma 3 12b (54.5%), Gemma 3 27b (60.3%). HiddenMath benchmark bar chart. Scores: Gemma 2 2b (1.8%), Gemma 2 9b (10.4%), Gemma 2 27b (14.8%), Gemma 3 1b (15.8%), Gemma 3 4b (43%), Gemma 3 12b (54.5%), Gemma 3 27b (60.3%).

HiddenMath

An internal holdout set of competition math problems.

MMMU benchmark bar chart for Gemma 3 models. Scores: Gemma 3 4b (48.8%), Gemma 3 12b (59.6%), Gemma 3 27b (64.9%). MMMU benchmark bar chart for Gemma 3 models. Scores: Gemma 3 4b (48.8%), Gemma 3 12b (59.6%), Gemma 3 27b (64.9%).

MMMU

Evaluates multimodal understanding and reasoning across various disciplines requiring college-level knowledge.

Gemma Quantization-Aware Training (QAT)

Gemma QAT dramatically reduces memory requirements while maintaining high quality. This lets you run powerful models like Gemma 3 27B locally on consumer-grade GPUs like an NVIDIA RTX 3090.

keyboard_arrow_right
Read more

What’s new in Gemma 3

Welcome to Gemma 3

Discover the latest advancements in Gemma, Google's family of lightweight, state-of-the-art open models.

Deepdive into Gemma 3

Hear how the Gemma research team unveil the architecture, design principles, and innovations behind Google's family of lightweight, state-of-the-art open models.

Building intelligent agents with Gemma 3

Explore the development of intelligent agents using Gemma models, with core components that facilitate agent creation, including capabilities for function calling, planning, and reasoning.

Gemma 3 architecture and design

Discover how with Gemma 3, we have tried to push many of the limits of what makes a model usable and practical.

A truly multilingual Gemma 3

Building multilingual AI applications is crucial for reaching global audiences, and varied language proficiency remains a top developer priority.

Download Gemma 3