“Smaller, more recent models like Qwen, with 27 billion total parameters and 2 billion active parameters, are now superior in capability to GPT-4.”