North Small Translate is open weights and does 50 languages well

CohereLabs released North Small Translate as an open-weights model. It has 25 billion active parameters — the rest are gated off — and 218 billion total.

It's built for machine translation, not general chat. The team tuned it across 50 languages, so it handles the ones that matter for la comunidad (Spanish, Portuguese, Tagalog, Mandarin) without needing a separate fine-tuned model.

Open weights means you can run it yourself, drop it into your pipeline, or serve it through vLLM without paying per-token. No vendor lock-in.

For folks building translation tools — the ones that help abuelos read the doctor's notes or that power the next wave of low-cost translation APIs — this is a practical release. 25B is small enough to fit on a single GPU and fast enough for real-time use.

Why this matters for us: open-weights translation means the tools we rely on stay affordable and don't get held hostage by big AI companies charging per use.