Cohere's New Translation Model Tackles Enterprise Language Barriers Across 50 Languages

Last week, enterprise AI firm Cohere introduced North Small Translate, an open-weight mixture-of-experts translation system capable of handling 50 languages. The model surpasses DeepL and Google Translate on WMT26 benchmarks, though commercial deployment requires a Model Vault license from Cohere's managed inference platform. For non-commercial work, developers may obtain the model weights under CC BY-NC 4.0 licensing.

The release extends Cohere's existing translation portfolio, which previously included Tiny Aya and Command A Translate. According to the company, North Small Translate demonstrates superior performance relative to open-weight models under 1 trillion parameters and competing API-based translation services across multiple evaluation metrics. Cohere frames the offering as part of a broader sovereign AI initiative, enabling organizations to maintain greater autonomy over model deployment locations and data handling practices.

Cohere co-founder Nick Frosst attributes the model's efficiency gains to its non-reasoning design, which relies on learned statistical patterns rather than step-by-step logical inference. This architectural choice reduces token consumption during operation.

Translation gaps persist across most global languages

We spent nine years scaling an architecture invented to fix translation, and machine translation is still broken for most of the world's languages. General-purpose models get you most of the way and then stop. The next phase of enterprise AI in this space is smaller, more specialized, and runs inside your own walls.

Nick Frosst, Cohere co-founder

On WMT26 benchmarks, Cohere reported North Small Translate achieving an All Languages score of 83.60, surpassing Qwen 3.5 397B A17B at 81.56, GLM 5.2 FP8 at 76.50, DeepL NextGen at 81.37, Gemma 4 31B at 79.46, and Google Translate at 68.20. The model employs a mixture-of-experts architecture with 218 billion total parameters, of which 25 billion remain active, yielding a smaller computational and memory footprint than alternative systems. Direct model comparisons in this category face limitations, as not all vendors publicly disclose parameter information.

Long-form translation and data control emerge as critical enterprise needs

Machine translation allows documents to be translated from one language to another automatically. With current solutions, long documents start to fall apart. Google Translate scores 21.3 on our long-context test, Gemma 4 31B 19.4; we score 48.9. That's [for example] a safety manual that reads fine on page one… and has drifted by page ten. The other risk is where the text goes. Once you push HR policies or regulated documents through a third-party API, that data has left your building, and necessarily that means your control over it is diminished.

Nick Frosst

Frosst highlighted the model's capability to maintain translation coherence across extended documents and handle high volumes of sensitive materials, addressing performance gaps that plague existing solutions. The system supports 32 high-resource languages alongside 18 additional languages.

North Small Translate incorporates translation-specific workflow features integrated directly into the model. These include structured translation support for Markdown and JSON formats, instruction-following capabilities for tone and formatting preferences, and terminology guides enabling custom vocabulary application during translation.

Multi-pass workflows balance quality and efficiency

North Small Translate works with a multi-pass workflow. The model translates, reviews its own output, finds errors, and fixes them – and this is the same loop we used in training. We ship both because standard is one pass and built for volume, while the agentic [version] spends more tokens for 84.36 against 83.60 on WMT26. That difference ends up being worth it when the document is a contract or a safety procedure, for instance, but in other cases you'd rather optimize for efficiency.

Nick Frosst

The model leverages the same foundational architecture as prior Cohere systems but achieves performance improvements through post-training enhancements, including reinforcement learning and datasets tailored specifically for translation tasks.

Model steerability drives language tone and formatting

Frosst noted that generative machine translation models across the marketplace deliver superior quality and steerability—the ability to guide tone, formatting, and other stylistic elements—compared with Neural Machine Translation systems commonly deployed commercially, though at substantially higher cost.

RWS, a language technology and AI solutions provider, partnered with Cohere during development. Contributions from RWS's Language Weaver research teams and language specialists informed the model's real-world translation capabilities.

Developers may access the model weights at no cost for non-commercial purposes across three quantization options. A Hugging Face Space and API interface are also available for users without access to required hardware infrastructure.

Source: The New Stack