Google Unveils Gemma 3n: A Multi-Modal Marvel

In a groundbreaking announcement, Google has introduced Gemma 3n, now available for preview on the innovative LiteRT Hugging Face community platform. This next-gen multimodal small language model is capable of processing text, images, videos, and audio, marking a significant leap in AI capabilities.

Power and Flexibility: Two Variants to Choose From

Gemma 3n comes in two powerful versions: the 2B and 4B models. Both support text and images, with audio functionality set to arrive soon. This is a monumental upgrade from the earlier non-multimodal Gemma 3 1B, which required just 529MB and was limited to processing 2,585 tokens per second using a mobile GPU.

Game-Changer for Enterprises

For enterprises, Gemma 3n opens up a realm of possibilities. With full device resources at their disposal, developers can leverage larger models right on mobile platforms. Imagine field technicians taking a photo of a malfunctioning part and instantly querying the model for clarity, or warehouse workers updating inventory via voice commands, all while their hands are occupied.

Efficiency Meets Innovation: Selective Parameter Activation

Gemma 3n employs selective parameter activation for efficient parameter management. This means that both the 2B and 4B models boast a greater number of parameters, optimizing performance during inference while still maintaining efficiency.

Enhanced Fine-Tuning Capabilities

Google is empowering developers to fine-tune the base model and subsequently convert and quantize it through advanced quantization tools available on Google AI Edge. The latest release includes next-generation quantization methods that boost model efficiency, reducing language model sizes by up to 4X while slashing latency and peak memory utilization.

The Power of RAG: Customization Like Never Before

Not just limited to fine-tuning, Gemma 3n can also be utilized for on-device Retrieval Augmented Generation (RAG). Currently available on Android, this feature enhances language models with application-specific data, allowing for an extremely customizable RAG pipeline, from data import to response generation.

Introducing AI Edge On-Device Function Calling SDK

In tandem with Gemma 3n, Google has rolled out the AI Edge On-device Function Calling SDK for Android. This innovative tool transforms LLMs from mere text generators to actionable entities that can execute real-world functions like setting alarms or making reservations, based on the user's inputs.

Seamless Integration of LLMs and Functions

Integrating an LLM with external functions has never been easier. Users simply describe the function's name, purpose, and required parameters through a Tool object passed to the large language model. This allows for receiving function calls from the LLM and returning execution results seamlessly.