“The best fully open language models are currently approaching the performance levels of the original GPT-4.”