Cohere ships North Small Translate as open weights, but locks out commercial use
Cohere releases North Small Translate 1.0, a 218-billion-parameter translation model whose weights are downloadable under CC BY-NC 4.0 but not usable in production without a commercial deal. The move confirms a shift: open weights no longer mean an open license.
September 11, 2026. Cohere is releasing North Small Translate 1.0, a machine translation model whose weights are available to download, but under a license that forbids commercial use. The move, announced under the banner of AI sovereignty, marks a turning point: for the most capable open-weight models, the word “open” is now negotiated at the edge of the license.
A translation model that beats DeepL and Google Translate
North Small Translate is a mixture-of-experts (MoE) model with 218 billion total parameters, 25 billion active, and a 16,000-token context window. It covers more than 50 languages and their regional variants. According to Cohere, it outscores DeepL and Google Translate on machine-translation benchmarks.
The distribution is deliberately multi-tier. The model is available today on the free tier through the Chat V2 API. For non-commercial use, the FP8 weights are published on Hugging Face under CC BY-NC 4.0 — the license that permits sharing and adaptation but forbids any commercial exploitation. To run it in production, you must buy a commercial license and deploy the model through Model Vault, Cohere’s fully managed inference platform.
The message is clear: you can study and evaluate the model freely, but you cannot build a product on it without going back to the vendor.
“Open weights” no longer means “open source”
The episode is not isolated, and that is what makes it structural. Last month, the Chinese lab Z.ai published the weights of its flagship GLM-5.3 on Hugging Face while tightening its usage terms. Where GLM-5.2 shipped under the permissive MIT license, GLM-5.3 adds requirements for certain commercial users: companies with aggregate revenue above $10 billion over twelve consecutive months must pass a security review before hosting the model or its derivatives commercially.
Cohere is part of the same movement, but with a different boundary: no revenue threshold, but a blanket commercial ban on the published weights. The paradox is all the more visible because Cohere had, in June, released its first coding model, North Mini Code, under Apache 2.0 — permissive, with no commercial restriction. The Canadian lab has stayed quiet about why it changed course.
The underlying tension deserves a name. Cohere builds its pitch on sovereignty: control over where the model runs and who sees the data, aimed at regulated industries. A commercial license preserves that promise of data locality. But it stops short of vendor independence: the enterprise keeps its data and infrastructure, but cannot fork the model, build a product on it, or keep running it if the terms change at renewal.
What “sovereign” means in practice
The sovereignty rhetoric covers two distinct promises. The first is data locality: the model runs on your infrastructure, and the text being translated never leaves your perimeter. For a law firm or a hospital, that is decisive — a contract or a medical record should not travel through a third party’s API. The second would be vendor independence: the ability to fork, adapt, and keep running the model no matter what happens to the vendor. It is precisely that second promise that CC BY-NC 4.0 withdraws.
The consequence is clear. An enterprise deploying North Small Translate in production through Model Vault is buying locality, not autonomy. If Cohere tightens terms at renewal, if pricing climbs, or if the team wants to leave the managed platform, it is left depending on a model it is not allowed to re-host. That asymmetry explains why a blanket commercial ban is, in practice, more restrictive than a revenue threshold like GLM-5.3’s: a $10 billion cutoff touches only a handful of players, while a commercial ban touches anyone selling a product, at any size.
The license, and nothing but the license
It is also worth crediting what the release genuinely provides. The FP8 weights on Hugging Face let you evaluate the model on your own corpora, measure its latency, quantify its errors, and compare its output against DeepL or Google Translate — a transparency proprietary vendors do not offer. For a team weighing three translation vendors, that ability to run an independent benchmark has direct negotiating value: it turns a sales claim into a reproducible measurement.
But that transparency stops at the edge of production. The published weights are in FP8, a quantization that may not match the managed service, and commercial fine-tuning, alternative hosting, and redistribution remain forbidden by the license. In other words, you can judge the model freely, but you cannot build on it. That is a distinction worth holding onto before reading “open weights” as a seal of openness.
Verdict
The North Small Translate and GLM-5.3 sequence draws an emerging rule: the more capable an open-weight model becomes, the more its vendor restricts commercial use. The era of “download it, it’s yours” is closing in favor of “download it, evaluate it, then negotiate.”
If you are doing research, evaluation, or prototyping, download the FP8 weights freely and benchmark North Small Translate against DeepL and Google Translate on your own corpora. CC BY-NC 4.0 covers you.
If you must integrate translation into production, treat this model as a licensed product, not an open-source component. Budget for Model Vault and a commercial agreement, or switch to a genuinely permissive model if vendor independence is your criterion. The real cost of sovereignty is not the license fee — it is the dependency it installs, dressed up as autonomy.