ModelGemini

Google launches multimodal EmbeddingGemma 2

Original title: EmbeddingGemma 2: an open, lightweight multimodal embedding model

Plain language

What it is: Google launched EmbeddingGemma 2, an open and lightweight model for creating mathematical representations of data. The model now supports text, code, images, video, and audio in a single shared space.
How it works: It runs directly on personal devices, allowing data processing without sending information to external servers. It is modular, meaning developers can load only the parts they need, such as text only, to save device memory.
Where it is available: The model is available for download on Hugging Face and Kaggle, and will soon be on Gemini Enterprise Agent Platform Model Garden. It is designed to run locally on consumer hardware.
Example: Finding a specific video clip using a voice memo, or searching through hours of audio recordings using a text query.

What it is: Google released EmbeddingGemma 2, a 740-million parameter multimodal embedding model built on the Gemma 4 architecture under an Apache 2.0 license. It maps text, code, images, video, and audio into a unified vector space.
How it works: It features a modular design requiring 270M parameters for text, with optional vision and audio encoders. It utilizes Matryoshka Representation Learning to dynamically truncate output vectors from 768 down to 128 dimensions, and includes an 8K token context window.
Where it is available: The model is available on Hugging Face and Kaggle, with upcoming availability on Gemini Enterprise Agent Platform Model Garden. It runs locally on edge devices and browsers using tools like LiteRT, MediaPipe, transformers, llama.cpp, and Ollama.
Example: Performing local codebase indexing, semantic code search, or building offline RAG pipelines when paired with Gemma 4.

The headline and explanation on this page are based on Google's post. The full details are in the original.

EmbeddingGemma 2: an open, lightweight multimodal embedding model