Language coverage and alternatives

[santhoshtr]

Hi,
NLLB is good start, however, there are many other opensource models that were released in last few years. Wikimedia Foundation has been providing a machine translation service based on a collection of such models(all free and opensource) and has coverage for 250+ languages. See https://translate.wmcloud.org/ and https://diff.wikimedia.org/2023/06/13/mint-supporting-underserved-languages-with-open-machine-translation/

I wonder if it is possible to bring these powerful models optimized for CPU to this app. Disclaimer: I am lead developer of that MT system at Wikimedia Foundation.

[niedev]

Hi, really cool project, I will definitely look at the source code.

As for the models used in MinT, I have already evaluated Opus-MT (I need to test better the quality and performance and decide if it is worth using it, given the greater complexity in managing languages) and I also implemented Madlad-400 3B during the tests (as quality it is superior to NLLB but it "goes crazy" more easily and with 8bit quantization it consumes too much RAM (4 GB), I will evaluate its use when 4bit quantization will be supported by OnnxRuntime), I don't know the other models so I will definitely check them.

Thanks for the suggestions.