We train Japanese/English models and they perform quite well for their sizes on translation (it is one of the core capabilities we eval): Shisa V2 - a shisa-ai Collection
Give the Nemo 12B version a try: shisa-ai/shisa-v2-mistral-nemo-12b · Hugging Face
The UnPhi 14B version is more capable and not much larger, but I noticed it has a tendency to add banal notes. 