Cohere announced North Small Translate as its first North-family model dedicated to machine translation. The 218B total / 25B active sparse MoE supports 16k input and output context across 50+ languages. In Cohere's WMT26 evaluation judged by GPT-5.6-Sol, it scores 83.60 (84.36 with an agentic multi-pass refinement workflow). Weights are on Hugging Face under CC BY-NC 4.0 for research; commercial deployment runs through RWS Language Weaver. Minimum hardware is one B200 or two H100 GPUs at W4A4 quantization.