Google DeepMind · Multimodal Embeddings

EmbeddingGemma 2

Released 6 October as the follow-up to EmbeddingGemma 300M (covered here in week 32 of 2025), now embedding text, images, audio and video in one space. The text model is 270M parameters; adding the optional 170M vision encoder and 300M audio encoder makes a 740M multimodal model. An 8K-token window holds up to 5.5 minutes of audio, 29 images or 58 video frames. Vectors are 768 dimensions and can be cut to 512, 256 or 128 through Matryoshka Representation Learning. Google reports MTEB Code rising to 78.68 from 68.76, vendor-reported, and says it keeps the original's multilingual strength. Apache 2.0.