Google - Gemma

Gemma: Google’s New Generation of Open Models for Responsible AI

Overview

• Family of lightweight, state-of-the-art open models
• Inspired by Gemini, with two sizes: Gemma 2B and Gemma 7B
• Released with pre-trained and instruction-tuned variants

Features

• Responsible Generative AI Toolkit for safer AI applications
• Toolchains for JAX, PyTorch, and TensorFlow (native Keras 3.0)
• Ready-to-use Colab, Kaggle notebooks, and integration with popular tools
• Deployable on laptops, workstations, Google Cloud, and Kubernetes Engine

Performance

• Best-in-class performance for sizes compared to other open models
• Directly runnable on developer machines
• Surpasses larger models on benchmarks while ensuring safety

Responsible Design

• Automated techniques filter sensitive data from training sets
• Extensive fine-tuning and RLHF align models with responsible behaviors
• Robust evaluations, including manual red-teaming and adversarial testing

Toolkit

• Safety classification, model debugging, and guidance tools
• Prioritize safe and responsible AI applications
• Access best practices for model builders

Optimization

• Fine-tune Gemma models for specific application needs
• Supports multi-framework tools, cross-device compatibility
• Optimized for NVIDIA GPUs, Google Cloud, and Vertex AI

Getting Started

• Explore Gemma at ai.google.dev/gemma
• Quickstart guides available

Future

• Ongoing expansion of Gemma model family
• Events and opportunities for developers and researchers