共有:

North Small Translate

It out-translated DeepL and Google Translate across 50 languages.

Cohere's new translation-specialist model, North Small Translate, is a 218-billion-parameter mixture-of-experts model that beat both commercial translation services and major open-weight models on the industry-standard WMT26 benchmark.

AI Navigate Editorial2026.09.127 min read

83.6 DeepL Google Translate Gemma 4 GLM 5.2 Mistral L3 North Small Translate WMT26 average score
01
The Model

Why build a translation-only
model, right now

A contrarian move at a time when general-purpose models keep getting better at everything, translation included.

Cohere announced North Small Translate, a translation-specialist model. Per the official post, North Small Translate, it's the first entry in Cohere's dedicated translation "North" family — text in, text out only, with none of the general chat or image-understanding features a chat model carries. The model card is published on Hugging Face: CohereLabs/North-Small-Translate-1.0.

General-purpose large language models can translate, but they weren't built for it specifically. North Small Translate goes the other way: it narrows its scope to 50-plus languages and points its parameters squarely at maximizing translation quality itself. The idea of a dedicated model isn't new on its own, but shipping one in 2026 — as general models keep improving fast — is the interesting part. This benchmark result draws a concrete, numeric line between where a general model is good enough and where only a specialist model gets you the last mile of accuracy.

Translation via a general LLMNorth Small Translate
One model handles chat, summarization, and translationPurpose-built for translation, text in and out only
Translation quality rides on overall model capabilityA dedicated design that outscores major services on WMT26
Fixing a mistranslation means rewriting the promptAn agentic variant that detects and fixes its own errors

02
Under the Hood

218B parameters, only 25B ever fire

A mixture-of-experts design that keeps compute cost down without giving up accuracy.

218B
total parameters
25B
active parameters per token
83.6
WMT26 score (agentic variant: 84.36)

Out of 128 "expert" sub-networks, only 8 fire for any given token. Total parameter count is 218 billion, but actual computation is scoped down to roughly 25 billion — a sparser, cheaper way to run than mobilizing the full model on every request. Input and output each support up to 16,000 tokens.

On the WMT26 benchmark, it scored 83.6 averaged across all languages, ahead of commercial services like DeepL and Google Translate as well as large open-weight models such as Gemma 4 31B, GLM 5.2, and Mistral Large 3. An agentic variant that also detects and corrects its own errors pushes that further, to 84.36.

03
In Practice

Who this changes things for

Developers building multilingual RAG

You can slot a dedicated translation model into a pipeline alongside a general-purpose retrieval and summarization model, lifting translation quality specifically without touching the rest of the stack.

Businesses going multilingual

Coverage of 50-plus languages makes it a candidate for localization across multiple markets — though commercial use still requires clearing licensing first.

Single-language personal users

If you already operate entirely in one language day to day, this brings you almost nothing. It's a model that only pays off once you're crossing languages.


There are places where a model built to do
one thing exceptionally well still beats one that does everything.


04
Caveats

Where the license and language coverage stop

North Small Translate is released under CC BY-NC 4.0, for research and non-commercial use only — anyone wanting to run it commercially has to go through Cohere's sales team first, so this isn't something every team can drop into production tomorrow. The benchmark win is also an average across the language set WMT26 actually covers; performance on languages outside that set, or on colloquial or heavily domain-specific text, still needs separate verification. Taking the benchmark number at face value isn't the safe move — testing it against the actual text your own use case involves is.

Worth sitting with, too, is what this says about specialist models as a category. Shipping a narrowly scoped model while general-purpose models keep getting better across the board is a bet that, in situations where quality has to come first, a specialist still wins. Whether other domains — summarization-only, code-only — see more of these purpose-built models follow will hinge on how North Small Translate actually performs once it's used in production.