Cohere released North Small Translate — an open translation model based on a Mixture-of-Experts architecture for working with 50 languages. According to MarkTechPost metadata, the model uses 25 billion of 218 billion parameters per token and achieved a score of 83,6 in the Cohere WMT26.
The release is positioned as a high-performance solution with low resource requirements, focused on speed and cost-efficient translation. The weights are available free for non-commercial use; commercial access is via Cohere Model Vault or RWS Language Weaver.
In the original package there is no full text of the study, description of testing methodology or independent verification of the result. Therefore the characteristics presented should be regarded as information from the metadata of the Cohere publication and its MarkTechPost retelling, and not as a fully verified assessment of the model quality.
editorial commentary
Why it matters
Likely practical consequence — interest from organizations needing open weights and translation in many languages with limited resources. The next visible signals will be full technical documentation and independent quality assessments. Significant uncertainty stems from the package containing only metadata, with no testing methodology or third-party results.