Blog

Eight Languages I Don't Speak, and the Gate That Keeps Me Honest

An open atlas with maps of many countries and languages
Photo by Kyle Glenn on Unsplash

Here is an awkward confession for a guy selling a multilingual autocomplete: I speak one language. English. That is it. If you write to me in Spanish, I will smile and nod and understand roughly half of it, and my grammar in every other language I claim to support is not something I would put in front of a native speaker.

So when I decided WriteAmp should handle languages other than English, I had a problem most autocomplete makers do not have to think about. I could not test any of it by feel. I could not type a German sentence and know whether the suggestion sounded like a person or a phrasebook. And that is exactly why the feature is built the way it is: not around my instincts, but around a gate that refuses to ship a language until a quality check says it is actually good.

The trap most apps fall into

Most autocomplete apps handle this differently. They ship an English model, bolt on a promise of "multilingual support," and when you type in French, the model does its best impression of a French person who learned French from an English textbook. It is not a bug. It is the cheapest thing that can be called multilingual.

The problem is that a wrong-language suggestion is worse than none. If I am writing to a client in German and the app offers me English words, I do not think "how thoughtful." I think the app is broken, and I turn it off. A silent English fallback does not feel like a feature. It feels like the product giving up on you.

The rule: no silent English fallback

So I made a rule for WriteAmp. When the app detects you are typing in a language it supports, it routes that field to the one model that was actually trained on many languages, and it does not offer English words in the middle of your sentence. When it detects a language it does not support yet, it stays quiet and tells you so, instead of guessing.

English runs on whichever engine you picked, mini, midi, or max. Every other supported language runs on a single multilingual model, the max tier, because that is the one model in the catalog that was built for this job. The first time you type in a non-English language, that model downloads, and after that everything works offline.

The cheap version
English with a promise
  • One English model tries to fake every language
  • Suggestions drift into English mid-sentence
  • You correct it so often you turn it off
The honest version
A dedicated multilingual model
  • Non-English routes to a model trained on many languages
  • No silent English fallback
  • Unproven languages stay quiet instead of faking it

What is actually enabled, honestly

Here is the current list, with the labels I use in the app, because I think fake confidence is worse than an honest label.

✓
Supported on the multilingual model: Spanish, French, German, Brazilian Portuguese, Japanese, Korean, and Hindi.
✓
Experimental: Arabic. It works on the multilingual model, but I do not call it supported until right-to-left and IME testing is further along.
✕
Detected but not enabled: Simplified Chinese. It failed my quality check, so it stays off until a better model clears it.

That last line is the one I am proudest of, weirdly. Shipping Simplified Chinese and calling it done would have been easy, and nobody would have caught me but the people it disappointed. Keeping it off, and telling you it is off, is the whole point of the rule.

Why the gate exists

Here is the part that did not come from a marketing brainstorm. When I started, I tried to build a clever detector for Hindi written in English letters, a way of typing that millions of people use. I shipped it, and it was not reliable enough. It misread real chat, it could not follow people switching languages mid-sentence, and it produced wrong-language suggestions. So I removed it. The feature does not exist anymore, because I would rather not offer something than offer it badly.

That experience is the reason the current version is built around a gate instead of enthusiasm. Every language in the supported list had to pass a quality probe against the actual model before it was enabled. I do not get to declare a language done because it feels close. It has to pass the check, and if it fails, it waits.

The honest note

The multilingual model is a bigger download than the English-only tiers, because eight languages is a lot of brain. It downloads once, on your Mac, only if you actually type in a language that needs it, and after that everything works offline. Your words are not going anywhere to be translated.

What it is like to use

You do not have to do anything. There is no settings screen where you declare your languages in advance, because your languages are whatever you type. If you write to your aunt in German in the morning and your client in English in the afternoon, WriteAmp follows along, and it will not offer you English words in the middle of the German note. The settings screen shows you exactly what is enabled and what is not, with the same labels you see above. I wrote about how the engines work if you want the mechanical details.

The ask

If you type in one of the supported languages, the download is free for 30 days, and I would genuinely like to know whether the suggestions sound like you. That is the test no quality probe can run, because I cannot tell a good German suggestion from a great one. You can. If your language is the one I labeled experimental or coming soon, I am sorry, and I am working on it. I would rather tell you that than fake it, and that rule is the whole story of this feature.


Sources

On macOS 26+ Macs, Apple Intelligence mode stays free even after the trial ends — you always keep a working path to suggestions.

Written by Amit Ashwini, who builds WriteAmp and runs its marketing. More: why the Tab key beats the chat box · mini, midi, and max compared · benchmark methodology.