What’s new in Gemma 3?

Gemma 3: Latest Generation of Open Models from Google

Key Concepts: Gemma 3, open models, parameter sizes (1B-27B), versatility, multi-modality (text, images, video), context window (128K tokens), function calling, structured output, fine-tuning, deployment, Gemmaverse.

Introduction

Gus Martins, a Product Manager on the Gemma team, introduces Gemma 3, the latest generation of open models from Google. It builds upon the success of previous Gemma models, which have seen over 100 million downloads and a community creating over 60,000 variants. Gemma 3 is designed for speed and efficiency, enabling developers to build and deploy responsible AI applications on workstations, laptops, and smartphones.

Gemma 3 Model Family and Accessibility

Gemma 3 is a family of models ranging in size from 1 billion to 27 billion parameters. Responding to community feedback, Google has included a 1 billion parameter version, making it more accessible for resource-constrained devices. This provides flexibility in choosing the right model size for specific projects, from lightweight mobile applications to large-scale services handling complex documents.

Core Technology and New Features

Gemma 3 is built from the same research and technology that powers Gemini 2.0 models. Key new features include:

• Versatility: Gemma 3 excels at a wide range of language tasks and supports over 140 languages, enabling applications to connect with a global audience.
• Multi-Modality: Gemma 3 can handle inputs from text, images, and videos, allowing for interactive and intelligent experiences. It can handle multi-turn conversations with images and tackle challenging math and coding problems.
• Increased Context Window: The context window has been significantly increased to 128,000 tokens, allowing Gemma 3 to process vast amounts of information (equivalent to an entire novel) and generate more coherent and insightful responses.
• Improved Function Calling and Structured Output: Gemma 3 offers improved function calling and structured output, making it easier to integrate with other tools and services for building intelligent agents.

Fine-Tuning and Customization

Gemma 3 is designed for fine-tuning, allowing users to adapt it to specific needs, specialize it for particular industries, improve its performance in specific languages, or tailor its output style. Fine-tuning can be done on Google Colab, Vertex AI, or on a user's own GPU.

Deployment and Integration

Gemma 3 is supported by popular frameworks like transformers, JAX, Keras, and Ollama. Google has partnered with industry leaders like NVIDIA, Hugging Face, and AMD to ensure speed, efficiency, and seamless integration.

Getting Started and Community

Users can try Gemma 3 through Google AI Studio, the Google Gen AI SDK, or directly from platforms like Kaggle, Vertex AI, or Hugging Face. It can also be downloaded and run locally. The Gemmaverse is a new page showcasing applications built by the Gemma community.

Conclusion

Gemma 3 represents a significant advancement in open models, offering increased versatility, multi-modality support, a larger context window, and improved function calling. Its accessibility and ease of deployment, combined with the ability to fine-tune and customize, make it a powerful tool for developers building a wide range of AI applications. The Gemmaverse highlights the impact of the Gemma community and the potential of Gemma 3.