Introducing Gemma 3: Advanced and Easy-to-Use AI Models

Google is launching Gemma 3, a new collection of smart, lightweight AI models. These models use the same technology as our Gemini 2.0 models and are designed to work fast on all kinds of devices like phones, laptops, and workstations. Gemma 3 comes in different sizes (1B, 4B, 12B, and 27B), so you can choose the right one based on your device and performance needs.

Today we launched Gemma 3, our most advanced and portable open models yet. This collection of lightweight models is designed to run fast, directly on devices like smartphones and laptops, to help devs create responsible AI apps at scale. Learn more ↓ https://t.co/yMNquc87mA

— Google (@Google) March 12, 2025

New Features of Gemma 3

Gemma 3 comes with powerful new features to help developers.

• Great Performance: Gemma 3 works faster and better than many other models, even on a single GPU or TPU (a type of computer chip).
• Supports Many Languages: It can work with more than 35 languages and understands over 140 languages, making it useful worldwide.
• Can Handle Text and Images: Gemma 3 can read and understand text, analyze images, and watch short videos, which opens up new possibilities for creating smart apps.
• Handles More Information: It can process huge amounts of data (up to 128,000 tokens) to deal with more complex tasks.
• Automates Tasks: Gemma 3 helps automate tasks by supporting function calls, so you can create efficient workflows.
• Better Performance with Smaller Size: The new version is smaller and needs less power, but it still works really well.

Safety Measures and ShieldGemma 2 for Responsible Use

We have focused on safety while building Gemma 3.

• Safe to Use: Gemma 3 has been carefully tested to make sure it’s safe and unlikely to be used for harmful purposes.
• ShieldGemma 2: Along with Gemma 3, we’re introducing ShieldGemma 2, a tool that checks images for harmful content like violence or explicit material. Developers can adjust this tool to fit their needs, ensuring safe and responsible AI use.

Works Well with Your Existing Tools

Gemma 3 is designed to easily fit into the tools developers already use.

• Works with Popular Tools: You can use Gemma 3 with tools like Hugging Face, JAX, PyTorch, and more, so you can work with what you’re already familiar with.
• Start Quickly: You can start experimenting with Gemma 3 right away on platforms like Google AI Studio, Kaggle, or Hugging Face.
• Customize and Fine-Tune: Developers can easily adjust Gemma 3 models to fit their specific needs by using its updated code and instructions.
• Flexible Deployment: Gemma 3 can be deployed on various platforms like Google Cloud or other services, making it easy to use on different systems.
• Optimized for Different Hardware: It’s also optimized to work well on NVIDIA GPUs, Google Cloud TPUs, and even AMD GPUs, ensuring top performance on many devices.

Expanding the Gemmaverse and Supporting Research

The Gemmaverse is a growing community of AI tools and models created by developers around the world. For example.

• AI Singapore’s SEA-LION v3 helps people across Southeast Asia communicate by breaking down language barriers.
• INSAIT’s BgGPT is a model focused on the Bulgarian language.