> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Cohere opens a translation model scoring 83.6 across 50 languages
- URL: https://globalfeed.ai/en/coheres-north-small-translate-a-218b-moe-translation-model-scoring-83-6-on-wmt26/
- Published: 2026-09-11T17:56:45.000Z
- Updated: 2026-09-11T17:56:45.000Z
- Description: Cohere and Cohere Labs released an open-weight translation model with 218B total and 25B active parameters, covering 50 languages and scoring 83.6 on the company's own WMT26 evaluation.
- Author: GlobalFeed Editor
- Tags: Cohere, translation, MoE, x-cohere, dil-en, denetim-gerekli, video

Cohere and Cohere Labs have released **North Small Translate**, an open-weight machine translation model built as a decoder-only sparse Mixture-of-Experts (MoE) Transformer. It has 218B total and 25B active parameters, with roughly 11.5% of weights active per token: compute tracks the 25B active figure, while memory must still hold all 218B. Coverage spans 50 languages, from Albanian to Vietnamese.

On Cohere's own WMT26 evaluation, the model averages 83.6 across all languages. Cohere says that result beats DeepL and Google Translate as well as open alternatives such as GLM 5.2 and Mistral Large 3\. These are the vendor's own runs with GPT-5.6-Sol acting as judge, so they should be read as vendor-reported until independent WMT26 numbers appear.

## Regional scores, throughput and cost

On EU languages the standard model scores 82.17 against Gemma 4 31B's 72.73\. South Asia goes the other way: 86.16 for North versus 88.04 for Gemma. An Agentic variant runs a multi-pass workflow that finds and fixes its own errors. In Cohere's scoring bands, 80 to 100 covers perfect output or minor errors only.

On identical hardware at low concurrency, Cohere reports 112 output tokens per second against 81 for Gemma 4 31B; at high concurrency the figures are 39 and 30, which the company describes as up to 1.4x higher throughput. The gap widens on long text: translating two book chapters in a single call, the model scores 48.9, while Google Translate scores 21.3 and Gemma 4 31B 19.4\. Quality was measured per paragraph with xCOMET-XL.

Cohere's cost chart puts the model at 80.1 for $0.000676 per task, averaging 661 tokens. Gemini 3.1 Pro Preview (high) is listed at $0.038928 per task, about 58 times more, with Qwen 3.5 397B A17B at $0.004525 and Command A+ at $0.005158.

The model is free on Cohere's Chat V2 API until rate limits, can be self-hosted non-commercially, or licensed for commercial use, with three production checkpoints published. It is the first translation model in the North family, following Tiny Aya and Command A Translate, and was built with RWS, whose Language Weaver scientists and language experts worked on output quality. Full details are in the [MarkTechPost report](https://www.marktechpost.com/2026/09/10/cohere-releases-north-small-translate-a-218b-moe-translation-model-that-scores-83-6-on-wmt26-across-50-languages/?ref=globalfeed.ai).