Cohere released North Small Translate, an open-weights translation model, but production use needs a paid Model Vault licence

Cohere released North Small Translate 1.0 on 10 September 2026, its first translation model in the North family. Weights are published on Hugging Face under CC BY-NC 4.0, which permits research and non-commercial use, and a separate commercial licence sold with Cohere's Model Vault is required to put the model into production. Cohere's own announcement post lays out the split, and The New Stack's coverage on 11 September focuses on the gap between "sovereign AI" branding and the licensing terms.

Cohere describes North Small Translate as a mixture-of-experts model with 218 billion total parameters, 25 billion active, a 16,000-token context window, and support for over 50 languages. The company reports minimum hardware of one B200 or two H100s at W4A4 quantisation, and lists FP8 quantised weights on Hugging Face. It is available today on Cohere's free tier through the Chat V2 API.

The vendor's own numbers

Cohere says North Small Translate scored 83.60 on the WMT26 all-languages benchmark, using GPT-5.6-Sol as the judge. The company reports its comparison numbers as 81.56 for Qwen 3.5 397B A17B, 76.50 for GLM 5.2 FP8, 81.37 for DeepL NextGen, 79.46 for Gemma 4 31B, and 68.20 for Google Translate. An "agentic" variant that can find and fix errors in its own translations scored 84.36. These are Cohere's own benchmark and Cohere's own judge model, so read them as the vendor's claim rather than an independent result.

The company also reports 1.4x higher output throughput than Gemma 4 31B on identical hardware in its own testing (112 versus 81 output tokens per second at low concurrency, 39 versus 30 at high concurrency), and a long-context evaluation score of 48.9 against Google Translate's 21.3.

What the licence actually gates

The pattern Cohere used here is the interesting piece, and The New Stack anchors its coverage on it. Non-commercial users, including researchers and developers evaluating the model, can pull the weights from Hugging Face under CC BY-NC 4.0 and run them locally. A team that wants to serve the same weights in production has to buy a commercial licence and deploy through Model Vault, Cohere's managed inference platform. Self-hosting the CC BY-NC weights on commercial GPUs, or wrapping them in a customer-facing API, is not allowed.

Cohere is not the only lab drawing this line. The New Stack points out that Z.ai's GLM-5.3 shipped last month with new commercial restrictions on companies with more than 10 billion dollars in trailing 12-month revenue, and a required Chinese security review for large-scale commercial hosting. GLM-5.2 had shipped under a permissive MIT licence. The trend in these announcements is the same shape: an open-weights headline, and a commercial licence attached at the door of production.

For an engineering team evaluating machine translation, the decision is now different in each of two paths. If the goal is a research evaluation or an internal prototype, the model is fully available and Cohere reports strong numbers on its own benchmark. If the goal is a customer-facing translation feature, the choice is between paying Cohere for Model Vault, negotiating a private deployment, or picking a model whose commercial terms match the deployment intent, such as Gemma 4 or a truly permissive Mistral release. A CC BY-NC line on the weights closes the option of hosting them yourself.

Source

Cohere's announcement, Introducing North Small Translate. The New Stack's coverage of the licensing move, Cohere says AI sovereignty matters, except it just gated its new translation model for commercial use.