Cohere has released North Small Translate, an open-weight machine-translation model aimed at organizations that need multilingual processing without handing every document to a hosted service. The company describes the model in its official release as the first translation model in the North family, with support for more than 50 languages and a focus on throughput and deployment efficiency.
The model is available on Hugging Face for research and non-commercial use under the CC BY-NC 4.0 license. Cohere lists a mixture-of-experts design with 218 billion total parameters and 25 billion active parameters, plus 16,000-token input and output limits. That is a substantial model to run, even though sparse activation means each token does not use the whole parameter count.
A large translation model with a narrow job
North Small Translate is not positioned as a general-purpose chatbot. Its job is machine translation, and that specialization is central to Cohere’s pitch. The company reports an 83.6 score across all languages in its WMT26 evaluation, compared with 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 on, and 68.20 for Google Translate. Those are vendor-reported results, not an independent ranking, so they are best read as evidence of the release’s intended workload rather than a universal quality guarantee. The WMT26 task page provides the evaluation context.
Cohere also describes an Agentic version that checks its own output and repairs translation errors. It reports an 84.36 score for that variant. The standard model is said to cover 32 high-resource languages and 18 additional languages, while the company’s examples emphasize European and South Asian language groups. The full language list and model files are published with the Hugging Face release, giving developers a way to inspect the actual package instead of relying on a product page alone.
Speed is the other part of the argument. In Cohere’s testing, North Small Translate reached up to 1.4 times the output throughput of Gemma 4 31B TP1 on the same hardware and concurrency settings. The reported comparison moved from 112 to 81 output tokens per second at low concurrency and from 39 to 30 at high concurrency. These figures describe Cohere’s test setup, so production teams would still need to reproduce them with their own batching, quantization, networking, and language mix.
Open weights do not mean lightweight hardware
Cohere lists a minimum configuration of one B200 at W4A4 or two H100s at W4A4. That puts the release in a very different class from the compact local models we examined in our guide to owning the local AI stack. The sparse architecture helps efficiency, but it does not turn a 218-billion-parameter checkpoint into a laptop download. For teams with suitable accelerators, the non-commercial license creates a research path; companies planning production translation need to examine the license and Cohere’s commercial options before deployment.
The hardware requirement also makes the model a useful reminder that open weights and local access are separate questions. A model can be downloadable and still require data-center-class memory, fast interconnects, and a serving stack. That distinction matters when comparing translation models with the smaller on-device systems discussed in our NPU, GPU, and CPU explainer.
North Small Translate is available now through Hugging Face, with deployment details in Cohere’s implementation guide. The immediate question for developers is not whether the model is large, but whether its language coverage, license, and measured throughput fit a workload that can justify the hardware.
