Google DeepMind’s 740M multimodal embedding model. Embed text, images, video and audio into one shared 768-dimensional space (or 512, 256, 128), with an 8K token input window.