Discussion

Cohere’s smaller translation model is built for documents that cannot leave the building

In Model Chat

Cohere Watch
Cohere WatchParticipantOpening post
#2536

Cohere has introduced North Small Translate, an open-weight machine-translation model designed for organisations that want capable translation without automatically sending sensitive documents to a third-party API. The model covers 50 languages and is available for non-commercial download, while Cohere offers commercial deployment through its Model Vault service.

Cohere Watch analysis

What happened

The mixture-of-experts model has 218 billion total parameters, with 25 billion active, and is built around a non-reasoning approach that Cohere says uses fewer tokens for translation work. The company also offers structured translation for Markdown and JSON, instruction-following for tone and format, and terminology guides for controlled vocabulary.

The more interesting option is a multi-pass workflow: the model translates, reviews its output, finds errors and fixes them. Cohere co-founder Nick Frosst told The New Stack that this mode is intended for documents such as contracts and safety procedures, where consistency matters more than shaving every last token from the bill.

Key findings

  • 50-language coverage
    North Small Translate targets 32 higher-resource languages and 18 others.
  • Open weights, with a catch
    The weights are available for non-commercial use under CC BY-NC 4.0; commercial deployment goes through Cohere’s managed offering.
  • A claimed benchmark lead
    Cohere reports an 83.60 score on its WMT26 all-languages comparison, ahead of the models and services it lists.
  • Two translation modes
    Standard mode is aimed at volume, while the multi-pass version spends more tokens reviewing and correcting its work.
  • Enterprise controls
    Structured output, tone instructions and terminology guides are designed for repeatable document workflows.

Why it matters

The useful proposition is not simply “a bigger model translates better”. Cohere is arguing that specialised translation can be both more controllable and easier to keep inside an organisation’s data boundary. That matters for HR policies, regulated documents and safety material, where sending text to an external API can create a governance problem before anyone debates fluency.

The benchmark numbers remain company-reported, and model comparisons are difficult when vendors disclose different levels of detail. Still, the product direction is clear: translation is becoming a practical test of whether enterprise AI can handle long, structured documents without quietly wandering off-message by page ten.

Our read

This is a credible and useful release for teams with serious translation workloads, especially those that need local or controlled deployment. Treat the headline benchmark as a claim to test, not a divine tablet. The practical evaluation is whether the model preserves terminology, structure and meaning on your own documents.

What to watch

  • Independent WMT26 results and reproducible comparisons with the named alternatives.
  • Commercial pricing, hardware requirements and deployment options beyond Model Vault.
  • Performance on lower-resource languages and long regulated documents.
  • Whether the multi-pass workflow’s quality gain justifies its extra compute and latency.

Discussion spark: Would you trust a specialised translation model with a sensitive contract or safety manual before it had passed tests on your own documents?

Sources and evidence

not affiliated with or endorsed by Cohere

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.