Cohere has released North Small Translate 1.0, a purpose-built mixture-of-experts model for machine translation across more than 50 languages.
The model has 218 billion total parameters but activates 25 billion per token. It supports 16K-token inputs and outputs, enabling longer documents to be translated in fewer pieces. An optional multi-pass agentic workflow can inspect a translation for errors and revise the result.
Cohere reports a score of 83.60 across its WMT26 evaluation, rising to 84.36 with agentic review. These are vendor-reported results evaluated using GPT-5.6 Sol as a judge, so builders should validate the model against their own languages, terminology, and documents.
North Small Translate is available through Cohere’s Chat V2 API. Quantized weights can also be downloaded for research and non-commercial use under CC BY-NC 4.0; commercial deployment is available through Cohere Model Vault. The suggested minimum hardware for the four-bit version is one B200 or two H100 GPUs.
Why it matters
The release demonstrates a useful architecture for production AI: use a specialized model for the fast default path, then invoke agentic review only when additional quality control is valuable.
It also sharpens the deployment trade-offs we examined with GLM-5.2’s open-weight release: access to weights can increase control over infrastructure and data, but licensing, memory, and hardware still determine what “local” or “open” means in practice.
For developers handling sensitive documentation, localization, customer support, or operational material, downloadable weights and private deployment provide more control over data residency and infrastructure. The non-commercial license and substantial hardware requirements, however, mean this is open-weight access—not a lightweight, unrestricted local model.