Cohere AI Updates: September 11, 2026

1. Cohere Released North Small Translate, a 218B Open-Weight Translation MoE

Cohere. North Small Translate is a decoder-only sparse mixture-of-experts model with 218 billion total parameters and 25 billion active per token, using 128 experts with eight activated per token plus shared experts on every token, and sliding-window attention at a 4096-token window mixed with global attention at a 3:1 ratio. It covers more than 50 languages with 16K-token input and output limits, and scores 83.60 on WMT26 all-languages, rising to 84.36 for an agentic multi-pass variant that finds and fixes its own errors. That puts it ahead of Qwen 3.5 397B at 81.56 and DeepL NextGen at 81.37, and on long-context evaluation it scores 48.9 against Google Translate’s 21.3. Weights are gated on Hugging Face under CC BY-NC 4.0 for research and non-commercial use, with commercial licensing through Model Vault and a free tier on the Chat V2 API. Quantized variants need at least one B200 or two H100s. It was built with language technology firm RWS. Source