Introducing Gemma: New State-of-the-Art Open Models for Developers
Learn about Gemma, the new generation of open models from Google that assist developers and researchers in building AI responsibly.
Gemma Open Models
Gemma is a family of lightweight, state-of-the-art open models developed by Google DeepMind and other teams across Google. These models are built from the same research and technology used to create the Gemini models, and they are designed to be used responsibly.
Inspired by the Latin word 'gemma', meaning 'precious stone', Gemma aims to provide developers and researchers with powerful tools for building AI. Along with the model weights, Google is also releasing tools to support innovation, collaboration, and responsible use of Gemma models.
Key Details about Gemma
Gemma is now available worldwide and comes in two sizes: Gemma 2B and Gemma 7B. Both sizes have pre-trained and instruction-tuned variants. In addition to the model weights, Google is providing a Responsible Generative AI Toolkit that offers guidance and essential tools for creating safer AI applications with Gemma.
Developers can use Gemma with major frameworks like JAX, PyTorch, and TensorFlow through native Keras 3.0. The availability of ready-to-use Colab and Kaggle notebooks, as well as integration with popular tools, makes it easy to get started with Gemma.
Furthermore, Gemma models can be deployed on various platforms such as laptops, workstations, Google Cloud, Vertex AI, and Google Kubernetes Engine (GKE). Google has also optimized Gemma for industry-leading performance across multiple AI hardware platforms, including NVIDIA GPUs and Google Cloud TPUs.
State-of-the-Art Performance and Responsible Design
Gemma models share technical and infrastructure components with Gemini, Google's largest and most capable AI model. This enables Gemma 2B and 7B to achieve best-in-class performance for their sizes compared to other open models. Despite their smaller sizes, Gemma models outperform significantly larger models on key benchmarks while ensuring safe and responsible outputs.
Google has prioritized responsible design in Gemma. They used automated techniques to filter out sensitive data from training sets and employed fine-tuning and reinforcement learning from human feedback to align models with responsible behaviors. Robust evaluations, including manual red-teaming and automated adversarial testing, were conducted to understand and mitigate risks associated with Gemma models.
