Cohere unveils north small translate: a 218b moe model achieving 83.6 on wmt26 across 50 languages

Cohere has launched North Small Translate, a sparse Mixture-of-Experts (MoE) machine translation model with 218B total parameters and 25B active parameters. The model supports 50 languages and scores 83.6 on the WMT26 evaluation, outperforming competitors like DeepL, Google Translate, GLM 5.2, and Mistral Large 3. It is deployable via Cohere’s API, free until rate limits, or available for self-hosting non-commercially or commercial licensing. The model, part of Cohere’s North family, is built with RWS and incorporates expertise from Language Weaver. It uses a decoder-only sparse MoE Transformer, with about 11.5% of weights active per token. In tests, it outperformed Gemma 4 31B in Europe and South Asia, with scores of 82.17 and 86.16 respectively. It also demonstrated higher throughput and better performance on long documents, scoring 48.9 on translating two book chapters compared to 21.3 for Google Translate and 19.4 for Gemma 4 31B. Cohere’s cost analysis shows the model is significantly cheaper than alternatives like Gemini 3.1 Pro Preview and Qwen 3.5 397B A17B. The model is available via Cohere’s Chat V2 API, with three checkpoints for self-hosting.