Cohere Releases North Small Translate to Enable Sovereign Enterprise AI

Cohere has released North Small Translate, an open-weight machine translation model designed to give organizations greater control over their data privacy. By allowing enterprises to run high-performance translations locally, the model supports the concept of "sovereign AI," ensuring sensitive information does not leave an organization's internal infrastructure.

Developed in partnership with RWS, the model leverages RWS's Language Weaver research and language experts. North Small Translate is built on a mixture-of-experts (MoE) architecture, featuring 218 billion total parameters with 25 billion active parameters. It supports 50 languages in total, comprising 32 "high-resource" languages and 18 others. The model builds upon Cohere's existing multilingual foundations established with the Tiny Aya and Command A Translate families.

Performance and Benchmarks

In technical evaluations, the model demonstrated strong capabilities in both general and specialized translation tasks. On the WMT26 All Languages benchmarks, North Small Translate achieved a score of 83.60. It also showed significant strength in handling extensive documents, scoring 48.9 on Cohere's internal long-context translation test, a critical metric for maintaining consistency across long-form text.

The model is positioned as a specialized tool for enterprise-grade efficiency and token optimization, prioritizing the specific requirements of high-volume translation over general-purpose reasoning capabilities.

The Push for Data Sovereignty

This release addresses a primary pain point for regulated industries that handle sensitive documents, such as safety manuals or HR policies. Cohere co-founder Nick Frosst highlighted the privacy risks inherent in current translation workflows, noting that pushing regulated documents through third-party APIs means data has left the building and control over that data is diminished.

By providing an open-weight model, Cohere enables companies to avoid the security trade-offs associated with cloud-based APIs. This shift is particularly vital for sectors where data residency and strict privacy compliance are legal requirements, allowing them to maintain the quality of technical translations without sacrificing security.

Availability and Outlook

North Small Translate is currently available for non-commercial use via Hugging Face under the CC BY-NC 4.0 license. As organizations continue to move away from centralized AI dependencies, the industry will be watching how this MoE approach scales across more diverse language pairs.

While the model marks a step forward in sovereign AI, the broader challenge of machine translation remains. Frosst noted that despite years of scaling architectures to fix the problem, machine translation remains "broken for most of the world’s languages," suggesting that further iterations will be needed to achieve true global parity.