Cohere releases North Small Translate, an open-weight translation model beating DeepL and Google Translate on WMT26

Cohere has released North Small Translate, an open-weight machine translation model the company says outperforms commercial services from DeepL and Google on a standard industry benchmark, while costing a fraction of a cent per task to run. The model was announced September 10, 2026 and is available now on Hugging Face.

What's new

Cohere describes the release as "North Small Translate, a mixture-of-experts machine translation model with strong performance across 50+ languages." On the WMT26 benchmark — a standard yearly evaluation for machine translation quality — the company reports that "North Small Translate achieves an 83.6 score across all languages, outperforming proprietary models like DeepL and Google Translate."

Cohere is also making a cost argument alongside the quality claim: for commercial deployment, it says the model delivers "a strong 80.1 score at just $0.000676 per task" compared to competing paid translation services that charge significantly more per request.

The model is being released in two tracks: it's "now available for research and non-commercial use under a CC BY-NC 4.0 license" on Hugging Face, with a separate commercial license path for enterprise deployment.

Context

The release follows Cohere's "North" family naming convention for its sovereign and enterprise-focused models, and lands the same week the company published a technical post on its megakernel serving engine for another model in the line, North Mini Code, claiming faster inference on H100 hardware. Machine translation has increasingly become a target for open-weight labs looking to undercut incumbent paid APIs like Google Translate and DeepL, which have dominated enterprise localization workflows for years on quality and language coverage rather than price.

A mixture-of-experts architecture — which activates only a subset of the model's parameters per request — lets Cohere claim both strong benchmark scores and low per-task cost, since only a fraction of the full model needs to run for any given translation.

Why it matters

If North Small Translate's benchmark numbers hold up independently, it puts real price pressure on the two companies that have effectively owned enterprise machine translation: Google Translate and DeepL. Open licensing for research and a separate low-cost commercial path also lowers the barrier for smaller companies and non-English-first markets to build translation into products without paying per-call fees to a US-based incumbent — reinforcing the broader 2026 trend of "sovereign AI" positioning that Cohere has leaned into across its North model line.