Cohere Labs posts 218B-param translation MoE for 50 languages

Cohere Labs has posted North-Small-Translate-1.0 to Hugging Face, a translation-specialized sparse Mixture-of-Experts model with 25 billion active parameters out of 218 billion total. The router activates 8 of 128 experts per token, and the model covers 50 languages, from English and Simplified Chinese down to Maltese, Punjabi and Icelandic.

The model card is blunt about what it was built for: "The model was post-trained specifically for translation quality." On WMT26, it reports a baseline score of 83.60 across all evaluated languages, rising to 84.36 with an agentic multi-pass translation workflow. No per-language numbers are published.

Deployment is documented at three precisions. BF16 needs 4 B200s or 8 H100s; FP8 halves that; the NVFP4 W4A16 build runs on a single B200 or two H100s. The license is CC BY-NC 4.0, with an acceptable-use addendum, and the terms are explicit: "If you are interested in commercial use, please contact Cohere's Sales team."