Cohere put a number on commoditized translation on September 10. Its new open-weight model, North Small Translate, scores 80.1 on the company’s quality measure while costing $0.000676 per task, which Cohere works out at an average of 661 tokens per job.
The comparison table is the argument. Cohere places Gemini 3.1 Pro Preview at $0.038928 per task in the same measure, and reports a lead over Gemma 4 31B in Europe at 82.2 against 73.9. In South Asia the two run level. Everywhere outside Europe that the company tested, it says DeepL NextGen finished behind.
Long documents are where the gap widens. Asked to translate two chapters of a book in a single call, the model scored 48.9, against 21.3 for Google Translate and 19.4 for Gemma 4 31B on Cohere’s evaluation.
Under the hood sits a decoder-only sparse mixture-of-experts Transformer. It carries 128 experts in total, activates eight per token, and alternates sliding-window attention with global passes in a 3 to 1 ratio, a design inherited from Command A. Cohere credits that layout with keeping long inputs affordable.
Scale reads as 218 billion parameters overall and 25 billion active per token. It is the first translation entry in the North family, covering 32 high-resource languages plus 18 more. The catch sits in the licence, which permits research and non-commercial use only. For buyers, that clause may matter more than any benchmark.
