Meta's latest efficient 70B model matching Llama 3.1 405B performance with improved multilingual and tool use.