A 17-billion-parameter mixture-of-experts instruction-tuned model from Meta with 16 experts, designed for efficient and scalable text and image understanding.