Many models work across many languages, but quality varies. A few practices help.
Prompt in the Target Language — or Not
You can write instructions in English and ask for output in another language, or write everything in the target language. Test both; results depend on the model and language.
Specify the Variety
Name the regional variety and conventions: Brazilian or European Portuguese, Simplified or Traditional Chinese, formal or informal address.
Provide Terminology
Give a glossary of product names, technical terms and phrases that must be used or left untranslated.
Beware Uneven Quality
Models generally perform best in widely represented languages. For less-resourced languages, expect more errors in grammar, facts and cultural nuance.
Tokens and Cost
Many languages use more tokens per word than English, increasing cost and reducing how much fits in the context window.
Cultural Adaptation
Examples, idioms, humour, dates, currencies and units may need localising, not just translating.
Check Quality
Have fluent speakers review outputs, especially for customer-facing, legal or medical content. Automated checks catch format errors but not tone or meaning.
Evaluate Per Language
If your application serves several languages, build test cases for each and track quality separately.