Cohere's new translation model prioritizes data sovereignty over reasoning shortcuts

Cohere has unveiled North Small Translate, a machine translation model designed around a straightforward principle: keeping sensitive corporate documents on premises while delivering better quality than most competing systems.

The open-weight model, which spans 50 languages and operates with 218 billion total parameters (25 billion active), relies on a mixture-of-experts architecture to minimize computational demands. Developers can access the weights for non-commercial purposes under a Creative Commons license, while organizations seeking commercial deployment can use Cohere's Model Vault service.

Why translation still lags for most languages#

Despite nearly a decade of progress, translation remains a stubborn problem for the vast majority of the world's languages. General-purpose AI models deliver passable results up to a point, then falter. "The next phase of enterprise AI in this space is smaller, more specialized, and runs inside your own walls," according to Cohere co-founder Nick Frosst.

North Small Translate deliberately sidesteps reasoning—the step-by-step logical processes that many newer models employ. Instead, it relies on learned statistical patterns, a choice that reduces token consumption and improves efficiency without sacrificing accuracy on benchmarks like WMT26.

Performance and long-document handling#

On WMT26 testing, North Small Translate achieved a score of 83.60, outpacing several competitors including Qwen, DeepL's NextGen, and Google Translate. A particular strength emerges in handling lengthy documents: while Google Translate scored 21.3 on Cohere's long-context test and Gemma 4 scored 19.4, North Small Translate reached 48.9—critical for translating multi-page safety manuals or compliance materials that must remain consistent across all pages.

This advantage matters because existing translation services often degrade in quality as documents grow. Pages early in a lengthy document read clearly; by the final pages, meaning can drift significantly.

Data control as a primary feature#

For enterprises handling sensitive materials, where the data goes matters as much as translation quality. When organizations route HR policies or regulated documents through third-party APIs, that content leaves their infrastructure. Cloud-based systems necessarily limit the organization's control over information access and retention.