On 10 September, a Canadian AI company quietly published a translation model that, by its own measurements, beats DeepL — the Cologne-based service that half of European business runs on. Cohere's North Small Translate scored 83.6 on the WMT26 benchmark against DeepL NextGen's 81.37, according to MarkTechPost, and it did so as an open-weight model anyone can download. If your business writes invoices in German, answers customers in French and markets itself in English, this is your corner of the AI world — and it just got interesting.
What actually happened
Cohere released North Small Translate on 10 September, built together with RWS, the language-technology firm behind many large localisation programmes. It is a mixture-of-experts model: 218 billion parameters in total, of which only 25 billion do the work on any given word, which keeps it relatively fast to run. It covers 50 languages — the full spread of EU languages included — and handles 16,000 tokens of text in and out, enough for a long contract chapter in one pass.
The headline numbers are bold. On the WMT26 translation benchmark, Cohere reports 83.6 for the standard model and 84.36 for an "agentic" variant that reviews and corrects its own output. DeepL NextGen scored 81.37 in the same comparison, Google Translate 68.2. On long documents the gap widens dramatically: 48.9 against Google Translate's 21.3, per the published results. One caveat belongs right next to those figures, and we'll get to it.
The weights are free to download — with a catch. The model sits on Hugging Face under a CC BY-NC licence, meaning research and personal use only. Commercial use runs through Cohere's paid API and Model Vault, or through RWS's Language Weaver platform. Cohere quotes roughly €0.0006 per average translation task on its commercial tier — pocket change next to general-purpose models pressed into translation duty.
Why a Canadian benchmark matters in Ghent and Lyon
Translation is the AI Europeans already use daily. Long before chatbots, the webshop owner in Lyon was pasting German customer emails into DeepL, and the accountant in Antwerp was translating Flemish engagement letters for a French client. It is the least glamorous, most load-bearing AI in European small business. So when a serious challenger claims the crown from Europe's own champion, the practical question isn't patriotism — it's whether your daily tool is about to get better or cheaper.
Competition here works in your favour either way. DeepL has faced remarkably little direct pressure at the quality end of the market; Google Translate competes on ubiquity, not polish. A credible rival with published scores forces everyone to improve, and translation pricing — already cheap — tends to fall further when open-weight alternatives circulate. The New Stack notes that Cohere is deliberately framing this as an alternative to routing text through big third-party APIs, which brings us to the part European businesses should actually care about.
The part worth stealing: how you translate is changing
Modern translation models take instructions. North Small Translate, like DeepL's newer offerings, follows terminology lists and style guidance: always render "algemene voorwaarden" as "terms and conditions", keep the tone formal, leave product names untouched. That is a different job than the sentence-by-sentence guessing of five years ago, and it is the difference between a translated website that reads like your company and one that reads like a machine.
Whole documents beat fragments. The long-document scores matter more than the headline ones. A model that holds sixteen pages in view keeps "the Supplier" as the same party from clause 1 to clause 40. If you still translate paragraph by paragraph in a free web tool, you are getting the worst version of what this technology can now do — whichever vendor you pick.
The data question almost nobody asks
Every pasted email is a data transfer. Cohere co-founder Nick Frosst put it bluntly to The New Stack: once you push HR policies or regulated documents through a third-party API, "that data has left your building." Under GDPR, customer emails, employee contracts and medical correspondence are personal data, and the free tier of any web translator rarely comes with a data-processing agreement. Paid tiers usually do — DeepL's business plans, for instance, and commercial API contracts generally. If your team translates client material daily on free accounts, that is a quiet compliance gap worth closing this month, whatever the Commission's enforcement priorities.
Self-hosting is the theoretical answer, not yours. The open weights mean a company can run this model entirely on its own machines — genuine data sovereignty. But the minimum hardware is one or two data-centre GPUs, an outlay in the tens of thousands of euros before electricity. That option is real for a hospital group or a bank; for a twelve-person firm, the realistic move is a paid tool with a proper agreement, not a server room.
When this isn't the right move
Don't switch tools over a benchmark. The scores above are vendor-reported and were judged by another AI model; independent WMT26 validation is still pending, as MarkTechPost points out. A two-point gap on an aggregate benchmark can vanish on your specific language pair — DeepL remains excellent on the EU languages that likely matter most to you. If your current tool produces text your German customers don't wince at, the switching cost buys you little.
The free weights are not free for you. The CC BY-NC licence excludes commercial use, so downloading the model to translate customer documents would breach it. And if translation is a few emails a week, none of this is worth an afternoon of your attention — the free tools you know are fine for gist-reading a supplier's message.
What to check this quarter
Make a one-page inventory. Where does translation actually happen in your business — website, contracts, support inbox, product sheets? Who does it, with which tool, on which account? Then two questions per line: is anything personal or confidential going through a free consumer tier, and would a glossary-following tool make the output sound more like you? That hour of stocktaking is worth more than any benchmark table, and it usually surfaces one upgrade that pays for itself in saved rework.
Translation is a narrow, concrete corner of AI — which is exactly why it's a good place to figure out how your business decides on these tools at all. If you'd like a structured look at where AI genuinely fits in your operations, and where it doesn't, Cresly's AI Readiness Scan walks through your workflows and gives you an honest starting map.