Google's EmbeddingGemma 2 is a small open model for search across text, images and audio

Google's EmbeddingGemma 2 turns text, images, audio and video into vectors in one shared space, so a single model can power search or retrieval over mixed files. It has 744M parameters in total, and you can load only the parts you need, down to 271M for text only. Google reports it beats many models two to twelve times its size on standard embedding benchmarks, with especially strong code search, though its English text scores dipped slightly against the first version. Quantised versions used under 600 MB of memory on a Pixel phone in Google's tests; these are Google's own results and not yet independently checked. On the Hugging Face paper page there is one human comment, from the person who submitted it, praising the small size. No debate yet. Sorted take: worth a test if you run search or RAG over PDFs, slides or call recordings.