Google releases EmbeddingGemma 2 with code, image, video and audio support

Google has released EmbeddingGemma 2, a compact open-weight multimodal embedding model built on Gemma 4 under the Apache 2.0 license. It maps text, code, images, video and audio into a unified 768-dimensional space and is optimized for on-device deployment on phones and laptops.

• Model scales from 270M parameters for text and code to 740M for full multimodal support
• Matryoshka Representation Learning allows truncating vectors from 768 to 128 dimensions, a sixfold memory cut
• At 128 dimensions it retains about 90% quality for text and ~75% for image, video and speech retrieval
• Available on Hugging Face and significantly outperforms EmbeddingGemma 1 on code search