4 ms·
That's just the base/stock/instruct model for general use case. There gotta be a finetune specialized in translation, right? Any recommendations for that? Plus
by jan_Sate 1y ago
That's just the base/stock/instruct model for general use case. There gotta be a finetune specialized in translation, right? Any recommendations for that?
Plus, mistral-small3.2 has too many parameters. Not all devices can run it fast. That probably isn't the exact translation model being used by Chrome.
- JimDabell 1y agoI haven’t tried it myself, but NLLB-200 has various sizes going down to 600M params: https://github.com/facebookresearch/fairseq/tree/nllb/ https://github.com/facebookresearch/fairseq/tree/nllb/ If running locally is too difficult, you can use llm to access hosted models too.