Cohere claims a WMT26 lead with a non-reasoning translator

Cohere co-founder Nick Frosst told The New Stack that his team spent nine years scaling an architecture invented to fix translation, and machine translation is still broken for most of the world's languages. His answer is North Small Translate, an open-weight mixture-of-experts model covering 50 languages that deliberately does no reasoning at all.

At a glance

• Cohere released the weights for 50 languages, 32 of them high-resource and 18 others, and built the model with RWS, whose Language Weaver research teams and language experts worked on it during development.
• In Cohere's own evaluation the model scores 83.60 on WMT26 All Languages against 68.20 for Google Translate, and 48.9 on a long-context test where Google Translate scores 21.3.
• The weights are noncommercial under CC BY-NC 4.0, commercial use runs through Cohere's managed Model Vault, and some rival comparisons rest on parameter counts that vendors never publish.

If you have not followed Cohere's translation work, North Small Translate sits on a lineage that includes the Tiny Aya and Command A Translate families. According to the WMT shared-task pages published by Statmt, Aya Expanse and Aya 101 were listed among suggested models for the constrained open-weights track, which caps a system's total parameters below 20B and requires weights to be released. Cohere models, in other words, were already circulating in the machine-translation benchmarking community before this one.

The model scores 83.60 on WMT26 All Languages, Google Translate 68.20

Cohere's reported evaluation puts North Small Translate at 83.60 on the WMT26 All Languages benchmark. The comparison set it gives: 81.56 for Qwen 3.5 397B A17B, 81.37 for DeepL NextGen, 79.46 for Gemma 4 31B (on), 76.50 for GLM 5.2 FP8 and 68.20 for Google Translate.

Long documents open a bigger gap. Frosst says Google Translate scores 21.3 on Cohere's long-context test and Gemma 4 31B scores 19.4, while North Small Translate scores 48.9. His example is a safety manual that reads fine on page one and has drifted by page ten.

Cohere also claims the model beats similarly sized open-weight models under 1T parameters, and API-based translation models, across various dimensions of machine translation on average. Every one of these figures comes from the company.

218 billion parameters in total, 25 billion active per token

The architecture is mixture-of-experts, and that is where the size numbers come apart. A MoE layer replaces the usual dense feed-forward block with a pool of expert networks plus a small router that learns which tokens go to which expert; according to Hugging Face's explainer, a token can be sent to more than one expert, inference is faster than a dense model of the same total size, and every expert still has to sit in memory. Think of a triage desk sending each patient to the right specialist: only two doctors see you, but the whole clinic is staffed.

On top of that, the model does not reason. Frosst says the efficiency comes from exactly that: it leans on learned statistical patterns rather than a step-by-step logic process, so it spends fewer tokens. Cohere points to the resulting compute and memory footprint as smaller than that of the models it lists.

The agentic version scores 84.36 by checking its own work

North Small Translate runs a multi-pass workflow, Frosst says: the model translates, reviews its own output, finds errors and fixes them, and that is the same loop Cohere used in training. Two versions ship. Why two? The standard one is a single pass built for volume; the agentic one spends more tokens and scores 84.36 on WMT26 against 83.60, which Frosst frames as worth paying for when the document is a contract or a safety procedure, and not worth it when you would rather optimize for efficiency.

The workflow features sit in the model rather than around it: structured translation of Markdown or JSON documents, instruction following for a recommended tone and format, and terminology guides that hand the model specific vocabulary to use. The architecture matches earlier Cohere models; the gains come from post-training, including reinforcement learning and new datasets aimed at translation.

Free weights for noncommercial use, Model Vault for everything else

Developers can download the weights under CC BY-NC 4.0 in three quantizations, free of charge for noncommercial use. There is a Hugging Face Space and an API for anyone without the hardware. Commercially licensed deployment goes through Model Vault, a Cohere-managed inference environment, and Cohere presents all of this as part of its sovereign AI strategy for organizations that want control over where models run and how data is handled.

That framing is what Frosst keeps returning to:

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.

On cost, he places generative translation models at the top for quality and steerability, suggesting tone and formatting, but says they typically run much more expensive than the neural machine translation systems used in most commercial work.

Cohere itself flags the weak spot in its own comparison table: some model-to-model comparisons in this space are not fully substantiable, because not every vendor discloses parameter counts, and the benchmark numbers are the company's reported evaluation rather than an independent run. For a model pitched on keeping regulated documents inside your own walls, it reads oddly to us that the commercial path is a Cohere-managed environment rather than the weights you just downloaded.

What Model Vault will cost

No price is given for commercially licensed deployment, and no figure is attached to the gap Frosst describes between generative models and NMT systems. The open questions from here are practical: what Model Vault costs per seat or per token, whether anyone outside Cohere reproduces the 83.60 and 48.9 figures, and how much extra spend the agentic pass actually adds for the 0.76 points it buys on WMT26.