Most translation pipelines break in the same place: long documents. A model that handles a paragraph well can fall apart when you feed it two chapters in one call, and the usual fix — chunking — adds latency, cost, and a new class of boundary errors. Cohere’s North Small Translate, released September 10, 2026, is aimed squarely at that gap.
What the model actually is
North Small Translate is a mixture-of-experts translation model with 218B total parameters and 25B active, 16k input and 16k output context, and text-only input and output. Cohere reports it supports 50+ languages and is available on Hugging Face under a CC BY-NC 4.0 license for research and non-commercial use, with several near-lossless quantizations. Minimum hardware is listed as 1× B200 or 2× H100s at W4A4.
On WMT26 benchmarks, Cohere reports an all-languages score of 83.60, ahead of Qwen 3.5 397B A17B (81.56), DeepL NextGen (81.37), Gemma 4 31B on (79.46), and Google Translate (68.20). An agentic variant that can find and fix translation errors scores 84.36. These are Cohere’s own numbers, judged with GPT-5.6-Sol, so treat them as a starting point for your own evaluation rather than a settled ranking.
Throughput and long-context behavior
The more useful claims for builders are about serving, not just quality. Cohere reports up to 1.4x higher output throughput than Gemma 4 31B TP1 on identical hardware and concurrency — 112 vs. 81 output tokens per second at low concurrency, and 39 vs. 30 at high concurrency. That is roughly 30–38% more tokens per second, which matters most when outputs are long.
On Cohere’s long-context evaluation, which measures translating two book chapters in a single call with per-paragraph xComet-XL scoring, North Small Translate scores 48.9 — more than double Google Translate (21.3) and Gemma 4 31B (19.4). The supplied material does not break this down by language or document type, so the practical question is whether your own documents hold up at that length.
Cost per task, not cost per token
Cohere frames efficiency as cost per task rather than per token. For commercial licensing, it cites an 80.1 score at $0.000676 per task using 661 tokens on average, against $0.038928 per task for Gemini 3.1 Pro Preview (high) at 5,762 tokens. Qwen 3.5 397B A17B and Cohere’s own Command A+ land at $0.004525 and $0.005158 per task. If you are comparing vendors, this is the number to replicate: run your own document mix through each model and measure total cost per finished translation, including retries.
The licensing tradeoff
The open weights are non-commercial. Enterprises that need commercial rights, security, and a managed localization platform get North Small Translate through RWS’s Language Weaver product, developed in partnership with RWS. That split is the real decision point: research and internal prototyping can start today on Hugging Face, but production deployment routes through a commercial agreement.
If you are weighing where a model runs and what data leaves your boundary, the same tradeoffs show up in zero data retention routing for AI APIs. For a translation pipeline, the question is whether you keep the weights in your own environment or accept a vendor-hosted path.
What to test first
Before committing, run three checks: your longest real documents in a single call, throughput at the concurrency you actually serve, and cost per finished task across your language mix. Cohere’s regional results show the model running even with Gemma 4 31B in South Asia (86.2 vs. 86.7) while leading in Europe (82.2 vs. 73.9), so per-language evaluation is not optional. The supplied material does not specify quality on low-resource languages outside the 50+ supported list, so verify coverage against your own locale set.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.