North Small Translate 1.0
modelYour notes
Cohere's open-weight translation model and the largest open Cohere MoE to date: the Command A Plus foundation of 218B total and 25B active parameters (128 experts, 8 active plus a shared expert), specialised for 50 languages with 16K input and output. The report describes difficulty sampling to select hard documents and a five-step protocol of supervised fine-tuning, DPO, and online reinforcement learning, with throughput prioritised through a non-reasoning base and optional agentic behaviour for quality gains. On WMT26 it scores 83.60 (84.36 in agentic mode) against DeepL 81.37, Qwen3.5-397B 81.56, and Gemma 4 31B 79.46. Released 10 September 2026 under CC-BY-NC-4.0 and served as north-small-translate-1-0; authors include Tom Kocmi, Phil Blunsom, and Aidan Gomez.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| WMT26 | 83.60 | — |