“The smallest Jemma models are multimodal and can understand audio, video, and image inputs.”