Intel / video

Cohere releases open-weights North Small Translate model

Source: video.twimg.com
Author: Cohere
Date: Published Sep 10, 2026

Context

North Small Translate is a translation model whose weights anyone can download, but the license limits who may use them. Cohere introduced it in a post that recalls the architecture was first proposed to improve Google Translate. Its release article describes a model, which routes each input through only part of the network. It has 218 billion parameters in total, 25 billion of them active at a time, a 16k input and output , and support for more than 50 languages. The weights are released under a CC BY-NC 4.0 license, which covers research and non-commercial use.

Cohere's headline result is a score of 83.60 on the WMT26 all-languages benchmark. On the same chart it lists DeepL NextGen at 81.37 and Google Translate at 68.20. These are Cohere's own figures, and the article notes the WMT26 scores were judged by another AI model, GPT-5.6-Sol, rather than by human raters. The launch video, read from an automatic transcript, makes a related claim: the model may not win in every region, but its quality falls off less from one language region to another than other models do. The video names Europe, Southeast Asia and East Asia as strong areas.

The video also lists business uses such as product localization, customer support and regulated-industry workflows, and Cohere's article, addressing enterprises evaluating commercial licenses, puts the cost at about $0.000676 per task, averaging 661 tokens. The downloadable weights stay non-commercial: the article says enterprises that need more than open-weight research access can get the model through RWS's Language Weaver product, but it does not give commercial license terms.

Terms in this piece · Glossary

transformer — The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.

Read the source video.twimg.com