Benchgen
Models/cohere-labs/

North Small Translate 1.0

DraftPublic

Model Details

North Small Translate 1.0

Org Pricing License Modality Released

Quick answer: North Small Translate 1.0 is Cohere Labs' open-weight, non-commercial-licensed translation specialist, built in partnership with RWS/Language Weaver. It's a 218B-parameter MoE model (25B active) supporting 50+ languages, released September 11, 2026.

What North Small Translate 1.0 Is

North Small Translate 1.0 is a machine-translation-specialist model rather than a general-purpose chat model — it has a comparatively short context window (16k input, 16k output) versus typical frontier LLMs, reflecting its narrower task focus. Weights are open for research/non-commercial use (CC BY-NC 4.0); enterprises need a commercial license from RWS/Language Weaver. It ships in full-precision, FP8, and W4A16 quantized variants and runs on as little as 1×B200 (W4A4) or 2×H100s (W4A4).

Specifications

OrganizationCohere Labs
LicenseCC BY-NC 4.0 (open weights, non-commercial)
Release dateSeptember 11, 2026
ModalityText
Knowledge cutoffNot disclosed

Public Benchmark Scores

No independently-tracked Benchgen benchmark currently matches Cohere's reported evaluation for this model. Cohere's own release reports 83.60 on an internal "WMT26 All Languages" evaluation (LLM-judged by GPT-5.6-Sol across 50+ languages) — this is a different task set and scoring methodology from Benchgen's existing WMT25 page (WMT's official General MT shared task, human/XCOMET-XXL-scored across 12 directions), so the two are not treated as equivalent.

Specs as reported in Cohere's own North Small Translate announcement (cohere.com/blog/north-small-translate, September 2026); not independently verified by Benchgen.