Google released EmbeddingGemma 2. It is an open multimodal embedding model. It maps text, images, audio, and video into one shared space. It runs on consumer hardware. It uses the Gemma 4 architecture. It has 740 million parameters. It supports modular encoders for text, vision, and audio. It allows dynamic truncation of output vectors for storage efficiency. It offers an 8K token context window. It improves code performance by 9.92 points on MTEB Code. It enables on-device RAG pipelines with privacy and low latency. It is available on Hugging Face, Kaggle, and LiteRT. It works with tools like MediaPipe, transformers, and Ollama.
What changed
EmbeddingGemma 1 supported text-only embeddings with a 2K token context window.
What it unlocks
Generating embeddings for text, images, audio, and video in one model on consumer hardware.
- 740M parameters
- 270M parameters for text-only
- 170M vision encoder
- 300M audio encoder
What you need to act on it
- on-device deployment
- commercially permissive Apache 2.0 license compliance
Sources