The Languages AI Doesn't Speak

Four gates between people and AI

The most capable AI ever built is, for most of humanity, behind a series of doors. None of the locks are technical.

• Identity. Create an account. Verify a phone number. Agree to terms.
• Banking. Put a card on file — which assumes a bank account. Roughly 1.4 billion adults worldwide don't have one.
• Country. Vendors publish supported-country lists. If you are outside one, the answer is no, regardless of your ability to pay.
• Language. And if you clear all three, you still have to speak English — or one of the few dozen languages a vendor decided to support.

Sats4AI was built to take those doors off their hinges. No account, no card, no country check: you pay per request with Bitcoin Lightning and you get the tool. That is the same argument Bitcoin makes about money, applied to AI — permission shouldn't be a prerequisite for capability.

The first three gates were straightforward. The language gate is the one that needed real work, because it isn't a policy choice — it is a genuine gap in what the models can do.

How many languages does AI actually speak?

Fewer than you would guess, and the vendors are refreshingly open about it if you read the documentation instead of the marketing.

Anthropic publishes per-language scores for 15 languages. In their own table, Claude Sonnet 4.5 scores Yoruba at 79.7% of its English performance, and Claude Haiku 4.5 drops to 52.7%. Yoruba has roughly 45 million speakers. ChatGPT ships 59 interface languages. Google Translate advertises around 249 in the browser, but the API that software can actually call is materially smaller — and it needs a Google Cloud project with billing attached.

So there is a long middle of the world's languages — tens of millions of speakers each — where the answer from frontier AI is either nothing at all, or fluent-sounding output that a native speaker recognises immediately as wrong.

What we added

A second translation engine, built on MADLAD-400 (Google's open, Apache-2.0 translation model), running on our own GPU. It covers 452 languages. Pick a mainstream language and you still get our fast language model, which is more fluent and cheaper. Pick something the mainstream models don't serve, and you get the specialist.

251 of those 452 are languages ChatGPT, Claude and Gemini do not support at all. That is the honest count, using a deliberately generous reading of what the frontier vendors cover — if there was any reasonable case they handle a language, we left it out of the number.

Of those 251, 29 clear our measured quality bar — scored against professional human reference translations, not vibes. Those are the ones we will actively stand behind. Sorted by speakers, that list starts:

• Bhojpuri — ~50 million speakers
• Maithili — ~34 million
• Egyptian Arabic — ~100 million
• Moroccan Arabic (Darija) — ~30 million
• Chhattisgarhi — ~16 million
• Central Atlas Tamazight — ~14 million
• Magahi — ~13 million
• Rundi — ~12 million

Plus Manipuri, Kashmiri, Awadhi, Shan, Kachin, Crimean Tatar, Quechua, Nuer, Sango, Fulfulde, Kabiyè, and European minority languages including Latgalian, Friulian, Ligurian, Silesian, Faroese, Sicilian, Lombard and Pangasinan.

The Arabic that people actually speak

This one deserves its own paragraph. Frontier models handle Modern Standard Arabic — the formal register of news broadcasts and official documents. They are far weaker at the dialects roughly 130 million people use in daily life. Egyptian Arabic and Moroccan Arabic are separate entries in our catalogue, with their own measured scores, not an afterthought folded into “Arabic”.

We publish the quality, including when it's bad

Here is the part most language lists leave out. Translation quality across 452 languages is genuinely uneven, and — awkwardly — the languages with the strongest case for existing here often have the least training data behind them. So we measured every one and publish the result:

• Good — scored 45+ on chrF++ against professional human reference translations.
• Fair — 32 to 45. Usable meaning, expect rough edges.
• Unverified — no public benchmark exists for this language, so we could not score it. Offered, but untested.
• Experimental — measured weak, or failed our structural checks.

The tier appears next to the language when you pick it, before you pay, and it comes back with the translation itself. Of the 452, 256 are unverified and 61 are experimental. We could have quietly advertised “452 languages” and left you to discover that. We would rather you spend 50 sats testing one sentence and decide for yourself.

Because for a lot of these languages, the honest comparison is not against a great translation. It is against nothing existing at all.

The other three gates, for completeness

A rare-language translation starts at 50 sats — a few cents — and you pay for the one request you made. No signup, no subscription, no card on file, no cloud project, no monthly minimum, no country check. Scan a Lightning invoice, get your text.

That matters most for exactly the people this post is about. A Bhojpuri or Hausa speaker is statistically far more likely to be outside the banking system that every major AI vendor requires as the price of entry. Serving someone's language while still demanding a Visa card would only move the door, not open it.

The same week: Kimi K3 became the largest open-weight model ever

Moonshot AI released Kimi K3 — 2.8 trillion parameters, a 1-million-token context window, weights free to download under a permissive licence. The most capable intelligence ever published openly is now something anyone can hold.

Which exposes the gap this post is about. Open weights solve ownership. They do not solve access. A model you can download still asks you to speak English to use it, and every hosted version of it still asks for an account and a card. If you write in Magahi, “the weights are free” changes nothing about your afternoon.

So the two halves belong together, and they landed in the same week: the intelligence became open, and the door became unlocked.

Ask the smartest open model a question in your own language

This part already works. Our multilingual-ask endpoint chains the two engines in a single paid request: your question goes from your language into English, a frontier model answers it, and the answer is translated back into your language. One payment, one call, no English required at any point.

It runs on Kimi K2.5 today — K3's direct predecessor — and the pipeline is deliberately model-agnostic, so Kimi K3 slots in as the brain with no rework the moment we add it to our model list. That is the next thing we ship.

Which means the destination is concrete rather than aspirational: a Bhojpuri, Magahi or Egyptian Arabic speaker asking the world's largest open-weight model a real question and reading the answer in their own language, for a few sats, with no account and no bank. Not a translation of an English AI product. The model itself, reachable directly.

Ask a question and get the answer in your language

Translation is the first half. The second is that these engines chain: your question travels from your language into English, a frontier model answers it, and the answer comes back in your language. One payment, one request. So a Bhojpuri speaker can ask a frontier model a question and read the reply in Bhojpuri, without knowing any English.

A note on our own site, since people ask: the Sats4AI interface itself is currently translated into 7 languages (English, Spanish, French, German, Portuguese, Italian and Chinese) and follows your browser setting. That is separate from the 452 languages the translation service covers — the site UI has not caught up with the engine yet.