Cohere released North Small Translate on September 10, an open-weight machine translation model the company says outperforms DeepL and Google Translate on a standard industry benchmark, expanding the AI lab's push beyond enterprise chatbots and into a market long dominated by dedicated translation services.
North Small Translate is a mixture-of-experts model with 218 billion total parameters, of which 25 billion are active for any given request, and Cohere says it supports more than 50 languages in both directions. On the WMT26 industry benchmark, Cohere's own announcement reports an all-languages score of 83.6, ahead of DeepL's NextGen system at 81.37 and well ahead of Google Translate at 68.20; an agentic variant of the model that can check and correct its own output scored 84.36.
The model is licensed under Creative Commons' noncommercial terms, meaning researchers and hobbyists can download its weights from Hugging Face and run it themselves at no cost, while businesses that want to deploy it commercially must license it through Cohere's partner RWS, whose Language Weaver platform serves translation customers across most of the Fortune 100.
Built for long documents, not just sentences
Cohere is drawing a specific contrast with general-purpose chatbots pressed into translation duty: the company says North Small Translate holds up on long-document translation tasks, scoring 48.9 on an internal benchmark against 21.3 for Google Translate and 19.4 for Google's open Gemma 4 model, and that its accuracy does not fall off in lower-resource languages the way some competing systems' does.
The model requires meaningful hardware to run at full precision — Cohere lists a minimum of one Nvidia B200 GPU or two H100 GPUs — putting local deployment out of reach for casual users, though the company is offering free, rate-limited access through its own API.
Cohere has increasingly focused on specialized, enterprise-oriented models rather than chasing general chatbot benchmarks against OpenAI and Anthropic, and independent write-ups of the release pointed to the RWS partnership as a sign the company is betting on localization firms as a distribution channel rather than building its own consumer-facing product.
The release gives researchers a free way to benchmark against DeepL and Google's commercial systems, and it will test whether Cohere's benchmark lead holds up once independent evaluators run the model on their own translation tasks rather than the vendor-selected comparisons in the announcement.