Google DeepMind has released EmbeddingGemma 2, an open-source model designed to convert text, code, images, video, and audio into a unified 768-dimensional vector space. This architecture allows for cross-modal comparisons, such as searching for video content using text queries.
The model features a modular design with 740 million parameters, split into 270 million for text, 170 million for vision, and 300 million for audio. Users can load only the necessary components to save resources; for instance, the text-only configuration requires approximately 191 MB of RAM, while the full multimodal version uses about 567 MB.
EmbeddingGemma 2 supports over 100 languages and offers a context window of 8192 tokens. It utilizes Matryoshka Representation Learning, enabling the 768-dimensional vectors to be truncated to 512, 256, or 128 dimensions to optimize storage and search speed.
The model is available on platforms such as Hugging Face and Kaggle. It is compatible with frameworks including MediaPipe, LiteRT, vLLM, and Ollama. Developers are advised to use bfloat16 or float32 precision, as float16 can cause numerical instability.
