Gemma Unveils Advanced Open Models

Responsible by Design

Responsibility is at the core of Gemma's design, guided by our AI Principles. To ensure the safety and reliability of our pre-trained models, we've employed automated techniques to remove personal information and sensitive data from our training sets. We've also extensively fine-tuned and used reinforcement learning from human feedback (RLHF) to ensure our instruction-tuned models align with responsible behaviors. To minimize risks, we've conducted thorough evaluations, including manual red-teaming, automated adversarial testing, and assessments of potential misuse. These efforts are detailed in our Model Card.¹

Together with Gemma, we're launching a new Responsible Generative AI Toolkit to support developers and researchers in building safe and responsible AI applications. This toolkit includes:

• Safety classification: A novel method for developing robust safety classifiers with limited examples.
• Debugging: A tool to help you understand and correct Gemma's behavior.
• Guidance: Access to best practices for model development, drawn from Google's experience with large language models.

Optimized Across Frameworks, Tools, and Hardware

You can tailor Gemma models to your specific needs, such as summarization or retrieval-augmented generation (RAG), by fine-tuning them with your data. Gemma is compatible with a broad range of tools and systems:

• Multi-framework tools: Use your preferred framework with reference implementations for inference and fine-tuning, including multi-framework Keras 3.0, native PyTorch, JAX, and Hugging Face Transformers.
• Cross-device compatibility: Gemma models work seamlessly across various devices, from laptops and desktops to IoT and mobile devices, as well as cloud platforms, making AI accessible to a wide audience.
• Cutting-edge hardware platforms: In partnership with NVIDIA, we've optimized Gemma for NVIDIA GPUs, ensuring top performance from data centers to local RTX AI PCs.
• Optimized for Google Cloud: Vertex AI offers a comprehensive MLOps toolkit with various tuning options and one-click deployment using built-in inference optimizations. For advanced customization, you can use fully-managed Vertex AI tools or self-managed GKE, deploying to cost-efficient infrastructure across GPU, TPU, and CPU.

Free Credits for Research and Development

Gemma is designed for the vibrant community of developers and researchers driving AI innovation. Start exploring Gemma today with free access on Kaggle, a free tier for Colab notebooks, and $300 in credits for first-time Google Cloud users. Researchers can also apply for up to $500,000 in Google Cloud credits to boost their projects.

Getting Started

To learn more about Gemma and find quickstart guides, visit ai.google.dev/gemma.

As we continue to expand the Gemma model family, we're excited to introduce new variants tailored for various applications. Keep an eye out for upcoming events and opportunities to engage, learn, and innovate with Gemma.

We can't wait to see what you'll create with Gemma!