What's New in the Gemmaverse: Gemma 3, Multilingual Models, and Ecosystem Growth

The Vision Behind Gemma

• 💡 Gemma is Google DeepMind's open family of AI models, designed to make AI research and development accessible to everyone.
• 🚀 The core mission is to solve intelligence collaboratively, ensuring that advancements benefit society responsibly.
• ✅ Responsibility is a key principle, focusing on both benefiting people and protecting them from potential harms.

Evolution of Gemma Models

• 🌟 Launched last year with 2B and 7B parameter sizes, Gemma has seen over 150 million downloads.
• 🖼️ PolyGemma 2 introduced vision capabilities, allowing models to process images.
• 🛡️ Shield Gemma acts as a safety classifier to enhance the responsibility of AI systems.
• 🚀 Gemma 3, released earlier this year, offers four sizes (1B, 4B, 12B, 27B) and is highly capable.
• 🌐 Gemma 3 is now multilingual, supporting over 140 languages, a significant expansion from its English-only predecessor.
• 📚 The context window has been dramatically increased, with models supporting up to 128K tokens, enabling longer conversations and more complex inputs.

Gemma 3N: On-Device AI

• 📱 Gemma 3N is optimized for on-device performance, requiring as little as 2GB of RAM.
• 🔊 It introduces audio understanding capabilities, allowing direct interaction with the model via voice.
• 🛠️ This architecture is shared with Gemini Nano and enables local processing without internet connectivity.

Building with Gemma

• 💻 AI Studio provides an easy entry point to test Gemma models, with features like image input and direct deployment to Google Cloud Run.
• 🐍 Gemma integrates seamlessly with popular open-source tools like Keras, Ollama, Hugging Face Transformers, LangChain, and LlamaIndex.
• ⚙️ Quantization techniques (like INT4) reduce model size and memory requirements, making powerful models accessible on standard hardware.
• 💡 Quantization-aware training (QAT) checkpoints help preserve model quality even after quantization.

The Gemma Ecosystem and Fine-Tuning

• 🌐 The Gemmaverse is a vibrant ecosystem with tens of thousands of variants and tools built by the community.
• 🔧 Fine-tuning allows developers to adapt Gemma models for specific languages, domains, or capabilities.
• 📈 Gemma models consistently rank high in benchmarks like LM Arena, demonstrating strong performance and approachability.
• 🌍 Community-driven fine-tuning has led to models excelling in languages like Korean, French, Spanish, and Japanese, as well as specialized applications like medical AI (MedGemma) and sign language translation (SignGemma).
• 🛡️ Shujma 2 is a safety classifier specialized for image content, enhancing system responsibility.