We started translating ProleWiki to french locally which would have taken 8 days of local inference on my computer, and then we found out Mistral AI offers a free API. You get 500,000 input tokens per minute, and a total of 2 billion token generation per month. All you need is to give them a phone number.

Basically with this kind of limits we could translate all of ProleWiki English (about 50 million tokens currently) to, well, every single language in existence or something like that. (the answer is 40 per month per account).

All of their models are freely accessible like this (they have thinking models, coding, etc for different tasks) but the counterpart is they are allowed to train their models on what you send. Which is fine with me lol here have some communist propaganda in your AI.

I can only speak for French but the translation quality is very, very good with the 3.2 Large model. Previously I could only run the 3.2 Small on my computer. The speed is also much faster than my computer and because it’s an API you can multithread (running more than one instance of the script at a time).

For example it’s able to figure out acronyms and automatically reorder them. Oh and we’re also translating our entire English-language library which contains over 1400 books. And just to preface - yes many of these are already available in French but it’s not a problem, down the line we can easily replace the machine translation with the existing ‘official’ editions. That’s the beauty of the internet, this would have been too much of a pain in print.

Of course language capabilities depend on a lot of things, but we may look at this for other language instances if the result is worth it. It requires a lot of testing with native speakers, changing the prompt one word at a time to test different outputs on the same chunk of text each time.

It does introduce some artefacts but this is very easily fixable with some regex to clean up the files which I think a lot of people just don’t bother to do but it’s literally so simple. For example the Large model likes to say “Here is the French translation of the provided wikitext:” sometimes, but you can very easily remove that in bulk. The small model didn’t have this quirk funnily enough.

Honestly I have to wonder how long this will keep going, this can’t be profitable for them. I don’t know how Mistral AI makes its money but they aim to be one of the big players in the sphere (and they’re based in France so western tech).

I made the script with crush + deepseek API and adapted it to be compatible with their API which is OpenAI format, and it works out of the box. I just run the script on my computer and they do the heavy lifting. And the more you use crush the more you know what you should have in your script or what you should ask crush/deepseek, like any tool. If I had to redo this script from scratch I think I could get it in just one prompt, maybe 2.

You could also use crush with their API, it’s just that the data will be used to train their AI. I also don’t know how good their model is at coding.

I was worried about potential censoring (“I’m sorry I can’t help with that let’s talk about something else”) but it seems APIs are generally more permissive than the web interfaces, probably because it would drive away enterprise clients.

Anyway, instead of running cycles for 8 days with my GPU this should take only 2 days, though we are spreading the workload between 4 people and have added the entire Library to this run.

Anyway if you wanted to try all of this stuff out you could begin with the free Mistral API. Open this link and it should prompt you to make an account, follow the steps then you can generate an API key here: https://console.mistral.ai/home?workspace_dialog=apiKeys.