A 17-billion-parameter mixture-of-experts instruction-tuned model from Meta with 128 experts, optimized for high-performance reasoning and coding.