Cohere Released an Open-Weight 218B Translation Model That Beats DeepL and Google Translate on WMT26
Hugging Face (Cohere Labs)·medium signal
North Small Translate is a mixture-of-experts model with 218B total and 25B active parameters, covering 50 languages. It scores 83.60 on WMT26, and an agentic multi-pass variant scores 84.36, against DeepL NextGen at 81.37 and Google Translate at 68.20. On book-length translation in a single call it scores 48.9, against Google's 21.3. The weights are on Hugging Face in BF16, FP8 and NVFP4 under CC BY-NC 4.0, and commercial use goes through Cohere or RWS.