Embedding Gemma 2: On-Device Multimodal Embeddings and Retrieval Demonstrations

This video introduces Embedding Gemma 2, an open lightweight multimodal embedding model that maps text, images, video, and audio into a unified embedding space. It explains the model design and demonstrates on-device retrieval workflows using Google AI Edge tools.

Model design and capabilities:
- Unified embedding space for text, images, video, audio, or mixtures
- Under 1B parameters, up to 740M modular design, 768-dimensional vectors, Matryoshka reduction to 128 dimensions, 8K context

Demos and on-device workflows:
- Google AI Edge Gallery for natural-language media search
- Video Moment Finder for locating moments within video
- AI Edge Foresight for audio-stream embedding, question detection, and local RAG with models like Gemma 4

Tuning and access:
- Domain-specific fine-tuning for legal contracts, medical imaging, or tech product catalogs
- Download model weights from Hugging Face and explore Gemma guide notebooks

The video is suitable for developers and ML engineers interested in on-device multimodal retrieval, privacy-preserving RAG, and compact embedding models. It provides an overview of the model's intended use cases and where to access its weights and examples.