Two different things get worse
When a model is "weaker in a smaller language", that actually means two independent problems, and they are worth separating, because their fixes differ too.
One is quantitative: the same sentence breaks into more tokens, so it costs more and takes up more room. This is precisely measurable and knowable in advance for every model. The other is qualitative: the model saw less text in that language, so it is shakier on grammar, on domain terms and on register. This varies enormously between models and only shows up when you measure.
The tokenizer tax, in numbers
A 2026 study covering the 24 EU languages compared six different tokenizers. The result is surprisingly orderly — it lines up by language family:
| Language family | Multiplier vs English |
|---|---|
| English | 1.00 |
| Romance | 1.5–1.7 |
| Germanic | 1.7–1.9 |
| Slavic | 2.2–2.5 |
| Uralic and Baltic | 2.7–3.0 |
Spanish comes in at 1.46, Hungarian at 2.71, Finnish at 2.99. The most expensive languages cost roughly three times what English does — for the same text.
What does this mean in practice?
Three things, and all three are felt immediately.
The bill. If a task costs a hundred dollars a month in English, the same task in a Uralic language is two to three hundred. Not because the model is more expensive there — same model, same price per token — but because there are more tokens in it.
The room. Your context window empties that much faster. What is a 300-page document in English leaves room for about 110 pages in Hungarian. With long material this limit arrives sooner than you would expect.
The speed. Output is produced token by token, so the answer also comes back more slowly. In a user-facing application that shows.
Quality: there is an official yardstick now
For a long time there was nothing to hold on to: the big comparisons measure in English, and about performance in a smaller language there were anecdotes at best.
In July 2026 EU MMLU was published, produced by the European Commission's Directorate-General for Translation: over a thousand questions, hand-translated into 16 EU languages, with around 250 translation students from 21 universities involved. It is the first yardstick that was not made by machine translation, so it genuinely measures performance in the language itself. Worth knowing about — but not sufficient, for the reason below.
The five-minute test of your own
A public yardstick measures what it measures. Not your texts. Five tasks, each surfacing a different kind of failure:
- Grammar. Hand over a sentence and ask for it in a different register, a different person, a different number. Agreement and inflection errors jump out immediately here.
- Domain terms. Hand over a professional paragraph from your own field. Watch whether it uses the established local term or calques the English.
- Register. Ask for a customer reply in a direct, informal tone. Many models switch automatically to stiff officialese in other languages.
- Formats. Ask for a date, an amount, an address and a phone number in the local form. The English pattern often bleeds through here.
- Local facts. Ask about a national statute or institution. You are not checking whether it knows — you are checking whether it invents something instead of saying it does not know.
How to score it
Not by impression. For all five tasks, write down in advance what the good answer is, then count how many came out right. Run across two or three models this is half an hour in total, and it tells you more than any ranking.
One important detail: run the same test on the model's cheap and expensive tiers. Performance in a smaller language typically varies far more between tiers than between vendors — so it may be that you do not need to switch vendor, only tier inside it.
What to do when it is weak
Four things, from cheapest to most expensive:
- Write the prompt in English, ask for the answer in the target language. The model understands the instruction most precisely in English; the output language can be stated separately. This alone often solves the problem.
- Give it a glossary. Four or five terms from your field: "we write these like this". This is the highest-return sentence in the whole prompt.
- Give it two examples. One good sample in the target language is worth ten instructions — that is what Few-Shot Examples is about.
- Put the important instruction at the end. With long material this matters as much as it does in English, except the text becomes "long" sooner.
What not to do
Do not fine-tune for this. Language weakness is rarely fixable by training, and it costs orders of magnitude more than the four steps above. Fine-Tuning, RAG or a Better Prompt? tells you where the boundary runs.
Do not translate back and forth. "I'll write it in English and have it translated" is a double loss: translation adds meaning drift on top of the stiffness. Ask for it directly in the target language, with the tricks above.
What to take away
There are two separate problems: the tokenizer tax and language quality. The first you know in advance — up to three times as many tokens, visible on the bill, in the context window and in the speed. The second has to be measured: five tasks, two models, half an hour. And if the result is weak, first switch the prompt to English and add a glossary — changing model comes after that.