Multilingual by default, ready for more
PT, ES, IT, EN at launch. The voice model, the prompts, and the evals, what changes between languages.
What 'multilingual' actually means
It's not a translation layer on top of an English worker. Each language is its own voice, its own prosody model, its own eval set, and its own list of things the worker is allowed to say.
PT-PT and PT-BR are two different workers, not one. So are ES-ES and ES-MX. The accents matter, the idioms matter, and the formality registers matter.
Evals are the hard part
We have 1,200 graded conversation turns per language and we re-run them on every model update. A turn that scores 4.6 in English might score 3.1 in Italian because of a single conjugation the model gets wrong under pressure.
Without the per-language evals, we'd ship regressions every week. With them, we catch the regressions before the customer does.
What's next
FR and DE are in private beta. NL and PL are on the roadmap for Q4. Each new language is roughly six weeks of work, mostly in eval curation rather than model fine-tuning. The model is the cheap part now.