It still does a much better job at translation than llama 2 70b even, at 6.7b params