The same model, the same question, a different language. What changes under the hood, and what to do about it.
Ask a model a question in Spanish and it answers in fluent Spanish. That fluency hides three differences from the English experience: the model read far less Spanish during training, it spends more tokens per sentence to represent it, and when Spanish varies by country it quietly picks one.
None of this makes the models bad at Spanish. The major open-weights models are strong in Spanish, French, German, Portuguese, and Chinese. It does mean a few habits are worth knowing, especially if you write in Spanish every day.
The text a model learns from is mostly English, because most of the web is. Spanish is among the best-represented other languages, and the gap on everyday tasks is small. It widens on the long tail: regional idioms, specialized vocabulary, recent slang. For languages with far less online text than Spanish it widens a lot. When a model stumbles in a language, it is usually a data problem, not a reasoning problem.
Models split text into tokens, and the tokenizer was tuned on mostly English text. Common English words are one token; Spanish words are more often split into pieces, and accents and longer words add more. The same paragraph in Spanish typically uses noticeably more tokens than in English, often a quarter to a half more depending on the model. Two consequences: less Spanish fits in a context window, and a Spanish conversation costs more to run. Newer tokenizers have narrowed the gap, not closed it.
Spanish differs by country in pronouns, verb forms, and vocabulary. A model trained on the whole internet has seen all of them and, without instructions, produces a blend that leans toward whatever was most common in its data. If you need Spain Spanish with vosotros, or Mexican usage, or a formal register, say so at the start. The model will follow, and it will hold the choice better if you remind it in a long conversation.
Reasoning models often do their visible working in English even when the question was in Spanish, then answer in Spanish. That is a training artifact, not a problem, and the answers are generally as good. It does mean the thinking trace is not always in the language you used.
Moving meaning between languages is close to the core of what these models learned, and they are excellent at it, including tone and register when you describe what you want. The weak spots are the same as elsewhere: idioms with no equivalent, very recent terms, and text where the wrong dialect would be noticed. Give the model the audience and the variety, and check the idioms yourself.
Write in whichever language you think in; the model will not be confused. Tell it the dialect and register once. Expect slightly shorter effective context in Spanish. If you are learning, ask it to correct rather than rewrite and to explain the rule, and treat its explanations as a starting point to check, not an authority.
It is like a well-read colleague who grew up reading in English and learned Spanish from books and the internet. Fluent, widely read, occasionally off on which word people actually use in Seville versus Bogotá, and happy to adopt whichever you ask for.
Berges AI is built by a language school that teaches Spanish and English, so how models handle both is something we test rather than assume. All the models we run work well in Spanish, and the interceptors and rules apply whichever language you write in. Tell it the variety you want, and it stays there.
Try Berges AIA little, and mostly on the edges. For everyday writing, explanation, and translation, the major models are strong in Spanish. The gap shows on regional idioms, specialized vocabulary, and recent terms. In languages with much less online text, the gap is larger.
Because Spanish uses more tokens for the same meaning. The context window is counted in tokens, so the same window holds less Spanish text than English text.
Yes. Say so at the start: "responde en español de España, con vosotros", or whichever variety you need. It will follow, and reminding it once in a long chat keeps it there.
It is a useful practice partner and a fast explainer, but it is not a teacher. It can be confidently wrong about a rule or a usage, and it will not notice the mistakes you make consistently unless you ask it to. Use it alongside instruction, not instead of it.