Canadian AI company Cohere released North Small Translate 1.0 this week under a CC BY-NC 4.0 license, making the model weights available for download and study but blocking production use without a paid commercial agreement. The company frames the launch as part of its mission to deliver sovereign AI — control over where models run and who accesses the data — for regulated industries. But the licensing terms mean enterprises can keep their data and infrastructure while still needing Cohere's permission to deploy the model commercially.
North Small Translate is a mixture-of-experts model designed for machine translation across more than 50 languages and regional variations. It contains 218 billion total parameters, 25 billion of which are active during inference, with a 16,000-token context window. The model is live on Cohere's no-cost tier via the Chat V2 API, and the FP8 weights sit on Hugging Face for anyone pursuing non-commercial projects. Enterprises wanting to put the model into production face different rules: they must buy a commercial license and run North Small Translate through Model Vault, Cohere's managed inference platform. The company hasn't explained why it chose to restrict commercial rights for this release.
According to the report, the open-weight approach offers researchers, developers, and enterprises "flexible ways to evaluate and deploy machine translation while retaining control over their data and infrastructure." That pitch appeals to organizations chasing sovereign AI. But the arrangement keeps enterprises dependent on Cohere — they can't fork the model, build products on top of it, or guarantee continued access if licensing terms shift at renewal. Cohere isn't alone in drawing sharper boundaries around open-weight releases. Last month, Chinese lab Z.ai published the weights for its GLM-5.3 model on Hugging Face but added commercial restrictions for companies with over $10 billion in aggregate revenue across 12 consecutive months, plus a mandatory security review for hosting. The previous version, GLM-5.3's predecessor, shipped under the permissive MIT license with no such strings attached.
The decision marks a shift for Cohere, which in June released North Mini Code under an Apache 2.0 license without commercial restrictions. That earlier coding model was pitched as a response to developers demanding the same sovereignty guarantees long required by regulated industries. The company has stayed silent on why North Small Translate took a different path. The licensing structure reflects a broader pattern of AI companies attaching tighter terms to increasingly powerful open-weight models, even as they market sovereignty and control to enterprise buyers. For organizations evaluating sovereign AI, the question isn't just whether they control the infrastructure — it's whether they control the terms under which the model remains available. Cohere's approach delivers the first without guaranteeing the second. Commercial gatekeeping on open weights creates an awkward tension: enterprises gain data sovereignty but lose strategic independence, a trade-off that may not satisfy buyers seeking genuine autonomy from vendor lock-in.