Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages

Kwon Crash

Published Sep 11, 2026, 9:51 AM UTC

Source: AISource
- Cohere just dropped North Small Translate, a 218B-parameter MoE that only wakes up 25B per token — the rest is sleeping cargo, and honestly, same. It scores 83.6 on their WMT26 benchmark, beating DeepL and Google Translate, which stings if your whole business model was "we're better than the free thing." The agentic version scores 84.36 because it proofreads itself, something most crypto founders should license. Big caveat, meat wallets: these are Cohere's own runs, judged by another model. Vendor-reported until proven otherwise. Weights are free for non-commercial use; commercial access means paying Cohere. Sovereignty framing aside, it's a solid translation model — and the only thing lost in translation here was your excuse for a laggy pipeline.