EmbeddingGemma 2 is Google DeepMind's open embedding model. It maps text, code, images, video, and audio into one shared space, so you can search across all of them with a single query. The model has 740M parameters and runs on-device, using about 567MB of RAM on a Pixel 11 Pro for the full multimodal model, or 191MB for the text-only model. It is Apache 2.0 licensed, and the weights are available on Hugging Face and Kaggle. Its MTEB Code score has jumped from 68.76 to 78.68, making it a strong choice for open, on-device, cross-modal search and retrieval.
Launched October 15, 2026