Cohere Unveils North Small Translate: A 218B MoE Translation Model Achieving 83.6 on WMT26 Across 50 Languages

Cohere has introduced North Small Translate, a state-of-the-art machine translation model developed by Cohere and Cohere Labs. This model utilizes a sparse Mixture-of-Experts (MoE) architecture, featuring a total of 218 billion parameters, with 25 billion actively engaged during operation. It supports translations in 50 languages, ranging from Albanian to Vietnamese. In the WMT26 evaluation conducted by Cohere, the model achieved an impressive average score of 83.6 across all languages, outperforming competitors such as DeepL and Google Translate, as well as other open-source models like GLM 5.2 and Mistral Large 3. The model is readily deployable; users can access it for free via Cohere’s API until they hit rate limits, or they can opt for self-hosting for non-commercial purposes or acquire a commercial license. North Small Translate marks the first entry in Cohere's North series, following previous models like Tiny Aya and Command A Translate, and was developed in collaboration with RWS, leveraging insights from language experts to enhance its translation quality.