Cohere North Small Translate: Open-Weight MT Across 50+ Languages

Coffee Summary

  • FACT: Cohere released North Small Translate on 10 Sep 2026: 218B total / 25B active MoE, 16K context, 50+ languages.
  • CLAIM: Cohere reports WMT26 all-languages 83.60 and agentic multi-pass 84.36.
  • FACT: Weights are on Hugging Face under CC BY-NC 4.0; commercial use needs a Cohere license.
  • FACT: Chat V2 id is north-small-translate-1-0; HF id is CohereLabs/North-Small-Translate-1.0.

What happened

Cohere positions North Small Translate as a dedicated translator in the North family. It is a sparse MoE decoder-only Transformer with 128 experts and 8 active per token. It supports instruction-following translation, post-editing, quality estimation, and optional agentic refinement. BF16, FP8, and NVFP4 checkpoints are published.

Why it matters

Teams choosing open translation models trade document quality against cost and control. Cohere’s pitch is sovereign-friendly open weights with API convenience. Vendor claims include 48.9 on a book evaluation versus Google Translate 21.3, plus up to 1.4x output-token throughput versus Gemma 4 31B. Validate results on your own language pairs.

What changed

  1. New open-weight MT specialist, not a general chat model.
  2. Dual access through Hugging Face weights and Chat V2.
  3. Agentic multi-pass is a quality-versus-latency dial.
  4. Commercial production beyond NC research requires a license.

AIImpish Take

Test a golden set across priority language pairs, compare single-pass with agentic mode, and stress long documents. Treat headline scores as hypotheses until your own pairs pass; reserve multi-pass for high-stakes documents.