Is there any free machine translation option better than Google Translate V1?

[angelfire76]

The current options that require an API require a paid account as far as I can see. Is there a free model that does a better job than GT V1?

[niksedk]

You can use Google Translate V1 with a free quota each day.
Another option is DeepL free API (requires enabling of billing). Is probably a tiny bit better than Google Translate.

Also, more and more "Chat GPT" like models are coming all the time. Many you can run locally via "LM Studio" or Ollama, which will be supported in next SE version. These models are slow (unless you have a nice gfx card), and the quality is probably yet worse then Google Translate.

Libre translate is also okay, but still not as good as GT.

[fideliochan]

There is DeepLx that you can use DeepL free api without having account.

[smalhotr85]

how much Google Translate V1 with a free quota each day?

[gericho]

I got excellent and relatively fast translations with nllb-serve and facebook/nllb-200-distilled-1.3B, there is also a 3.3B version, LLM like Hermes 2 Pro and openhermes, are more colloquial language aware than NLLB, but worse in grammar. From Spanish to Sinhala, the readers refer the translation is almost perfect, way more grammar-corrected compared to Google V1! I'm not sure it's an exception with this language combo, but I advice anyone with the necessary hardware requirements to try it and give a feedback. Thank you!

EDIT: Actually, after some more testing, Google V1 is still the king, NLLB come second. LLMs I tried are still giving garbage output with these language pairs, BUT still very good for Latin-only base languages

[dgoryeo]

@gericho, thanks for the pointer. Would you have a sample code on or a notebook?

[gericho]

I checked today with mother-tongue people, they say 3.3B is slightly better than 1.3B. Be aware 3.3B would eventually run only on the CPU so painfully slow, whereas the 1.3B with 8GB VRAM runs way faster, not on par with an online service like Google V1 but you can get 2h translation in 5 minutes or so. The Google V1 btw is still superior to NLLB but only on some specific parts of the text. Keep in mind I can talk only for the language pair just mentioned above, so your mileage may vary.
About the guide, I just followed the Subtitle Edit FAQ in the NLLB section here

Just a side note, I think there will be more chances in using an LLM Studio approach and a properly trained model for a specific language pair from Huggingface for example, unfortunately, fine-tuned models are only a few.

[gericho]

I'd like to bring your attention to a very well-made and up-to-date model comparison by Wolfram Ravenwolf

I am testing the top-rated models and see how they perform.

[Fuckingnameless]

for brazilian portuguese the only viable option is ollama, google api's i'm not sure how to setup but if it's anything like translate.google.com it's worthless

[AuroraMartell]

for brazilian portuguese the only viable option is ollama, google api's i'm not sure how to setup but if it's anything like translate.google.com it's worthless

I use ChatGPT or Gemini for translation.