If I tell you the model will serve users in 40 languages, what changes?

Serving 40 languages requires a more robust tokenization strategy. If the tokenizer is trained on an English-dominant corpus, non-Latin scripts will suffer from high ‘fertility’ (too many tokens per word), hurting performance. A larger vocabulary (100k-200k) is necessary to accommodate the diverse character sets and linguistic structures, and the training data must be carefully balanced to ensure fair representation.