Ollama adds Google's EmbeddingGemma 2 for local multimodal search

Ollama has added EmbeddingGemma 2, Google's new model for representing different kinds of content in a form that software can search and compare. Google introduced the model on October 6 for applications running on consumer devices.

The complete model has 740 million parameters and builds on Gemma 4. Its design combines a text component with optional vision and audio components, allowing applications to load only the capabilities they need. Google released it under the Apache 2.0 license.

EmbeddingGemma 2 extends its predecessor's text focus to additional media, including images, recordings and video. Its shared numerical representations support searches across content types, such as finding an image from a written query. Google positions the model for local retrieval, classification and search applications.

Developers can download the model using the command ollama pull embeddinggemma-2. Ollama's version 0.40.0 release notes list support through MLX, its runtime option for compatible models on Apple Silicon.