Machine Translation Is Still Broken for Most of the World's Languages: Cohere Builds Non-Reasoning for a Reason
The New Stack, Sunday, September 13th, 2026
The New Stack examines Cohere's North Small Translate, a 218B-parameter MoE open-weight translation model built deliberately non-reasoning.
The New Stack digs into Cohere's North Small Translate, a mixture-of-experts open-weight machine translation model covering 50 languages, with 218 billion total parameters and 25 billion active.
Co-founder Nick Frosst argues the model is deliberately non-reasoning: it leans on learned statistical patterns rather than step-by-step logic, which cuts token use and compute.
Cohere's reported WMT26 all-languages score of 83.60 puts it ahead of Qwen 3.5, DeepL NextGen, GLM 5.2 FP8, Gemma 4 and Google Translate. Weights are CC BY-NC 4.0 for noncommercial use, with commercial deployment via Model Vault.