Google EmbeddingGemma 2 adds images, audio, video

Google released EmbeddingGemma 2, an open multimodal embedding model that extends the original text-only version to handle images, audio and video within a single embedding space. Built on the Gemma 4 architecture, it has 740 million parameters, more than twice the original. The text-only core uses about 191 megabytes on a Pixel 11 Pro. Code retrieval scores jumped nearly 10 points on the Massive Text Embedding Benchmark. Weights are available on Hugging Face and Kaggle under an Apache 2.0 license, and the model works with vLLM, llama.cpp and Ollama.