I haven’t tried it myself, but NLLB-200 has various sizes going down to 600M params:

https://github.com/facebookresearch/fairseq/tree/nllb/

If running locally is too difficult, you can use llm to access hosted models too.