Cohere spent the last year building a 218-billion-parameter mixture-of-experts model. It cannot write code. It cannot hold a conversation. It cannot look at an image. It does exactly one thing: it turns text in one language into text in another, and according to Cohere, it does that better than anything DeepL or Google currently sells.
North-Small-Translate-1.0 landed on Hugging Face on September 10, 2026, under a CC BY-NC 4.0 license — open weights, non-commercial use only. It is the first translation model in Cohere's North family, and it is architecturally serious: a decoder-only sparse MoE Transformer with 218B total parameters and 25B active, routing each token through 8 of 128 experts plus a shared expert applied to every token. Context runs 16K tokens in and 16K out. Coverage is 50 languages — Cohere's announcement says 50+, though the model card enumerates exactly 50, a small gap worth noting in a launch otherwise built on precision.
The headline number is 83.60 on what Cohere calls the WMT26 all-languages benchmark. An agentic multi-pass variant, which the company says can find and fix its own errors, reaches 84.36. In the same evaluation, Cohere placed Qwen 3.5 397B A17B at 81.56, DeepL NextGen at 81.37, Gemma 4 31B at 79.46, GLM 5.2 FP8 at 76.50, and Google Translate at 68.20 — a 15-point gap over the most widely used translation system on earth.
Where the specialization shows most is length. On Cohere's long-context test, which measures translating two chapters of a book in a single call with per-paragraph quality scored via xComet-XL, North Small Translate scored 48.9 against 21.3 for Google Translate and 19.4 for Gemma 4 31B. Long documents are where general-purpose models drift, lose terminology consistency, and quietly paraphrase. Cohere is arguing that a model trained for one job does not.
The RWS story behind the model
The model did not emerge from a lab in isolation. It was built alongside RWS, the localization giant whose Language Weaver platform sits inside a large share of regulated enterprise translation workflows. The collaboration started in July 2025, when Cohere asked RWS to stress-test Command A Translate before launch. By September 2025 the two teams were building something purpose-made.
RWS CEO Ben Faes has been unusually direct about how deep it went. “The word 'partnership' understates the relationship that we've built with Cohere,” he said, describing the resulting model as “the brain behind the product.” On the earlier Language Weaver Pro system that came out of the same collaboration, Faes was blunter about what he thinks general-purpose models get wrong: “Most AI translation speaks the language but misses the meaning.”
That earlier model, Cohere reported, won 55% of sentence-level head-to-head tests against DeepL NextGen in human evaluation and beat competitors in automated benchmarks across 31 of 32 enterprise languages. “The test results have been phenomenal,” Faes said. “I'm amazed that we're getting to that level of quality in such a short time.” RWS says it works with more than 80% of the world's top 100 brands — which is the actual distribution channel here, and arguably the point.
Why this matters
Read the fine print on the benchmarks. These are vendor-reported scores, not results from the official WMT shared task, and WMT has not independently confirmed them. Cohere ran its own evaluation on WMT26 data using GPT-5.6-Sol as the judge — an LLM-as-judge setup that is now common practice and still carries known failure modes, including a documented tendency for model judges to reward fluent, LLM-shaped output over terse, literal, professionally correct translation. That bias, if present, points in exactly the direction that flatters a model like this one against a system like Google Translate. There is also a smaller inconsistency worth flagging: Cohere's own cost-efficiency chart reports North Small Translate at 80.1, not 83.60, under a different configuration. None of this means the model is bad. It means the margin is not yet an independent fact.
The strategic read is more interesting than the scores. For three years the industry's working assumption was that frontier generalists would absorb every narrow task — that a model good at everything would eventually be good enough at translation to make dedicated systems obsolete. Cohere is betting the opposite, and pricing it accordingly: it claims $0.000676 per task at 661 average tokens, against $0.038928 for Gemini 3.1 Pro Preview. At enterprise localization volumes, a 57x cost difference is not a rounding error. It is the entire procurement decision.
The license is the tell. Free weights, free API access, and no commercial use — that combination is not open-source idealism, it is a funnel. Researchers and evaluators get the model for nothing; anyone who wants to run it in production goes through Cohere's Model Vault or RWS Language Weaver. Cohere has spent two years positioning itself as the sovereign-AI vendor for regulated industries, and translation is the ideal wedge: high volume, high compliance burden, and currently split between two American providers that many European, Canadian, and public-sector buyers would rather not depend on. Minimum hardware is one B200 or two H100s at W4A4, which makes on-prem deployment genuinely feasible rather than theoretical.
What to watch: whether independent WMT26 results reproduce the margin over DeepL, whether the agentic variant's extra 0.76 points survives contact with real latency budgets, and whether a competitor answers with a specialized model of its own. If Cohere's numbers hold up under outside evaluation, the more consequential story is not that Cohere beat DeepL. It is that the era of one model doing everything may have just found its first real exception.
“Most AI translation speaks the language but misses the meaning.”— Ben Faes, CEO, RWS