EmbeddingGemma 2: Unified and Efficient Multimodal Representations

EmbeddingGemma 2 is an open model designed to map text, images, audio, video, and combinations of these into a shared embedding space. Such representations can support tasks including search and matching across different media types. The work aims to combine broad multimodal coverage with strong performance, rather than relying on separate specialist models for each kind of input. Its design includes simplified per-layer embeddings and a staged training curriculum that progressively introduces modalities and learning objectives, intended to maintain quality on single-modality, cross-modal, and mixed-input evaluations.

The authors report that the model sets a new Pareto frontier on comprehensive embedding benchmarks spanning all supported modalities, outperforming specialist systems as well as models two to twelve times larger. They also emphasize deployment flexibility: users can load only the modality encoders they need, adjust the number of vision tokens, shorten embedding dimensions, and use quantized weights to fit tighter compute and memory budgets. The paper analyzes key design choices through ablation studies and releases the model to support further research. The combination of multimodal breadth, benchmark results, and controls for adapting resource use is central to its case for a unified embedding system.