Cohere · 2026-09-10 · major
North Small Translate — Cohere's open translation model beats DeepL on WMT26
North Small Translate is a 218B-parameter open-weights translation model from Cohere Labs with 25B active parameters. It scores 83.60 on WMT26 across all languages, ahead of DeepL NextGen at 81.37.
Cohere Labs opens the weights of a 218B mixture-of-experts model built only for translation across 50 languages.
Quick facts
| Maker | Cohere Labs |
|---|---|
| Parameters | 218B total, 25B active (MoE) |
| Languages | 50 |
| Context window | 16K in / 16K out |
| License | CC BY-NC 4.0 (non-commercial) |
| API model ID | north-small-translate-1-0 |
| Commercial access | RWS Language Weaver |
Benchmarks
| North Small Translate (agentic) | 84.36 | |
|---|---|---|
| North Small Translate | 83.6 | |
| Qwen 3.5 397B A17B | 81.56 | |
| DeepL NextGen | 81.37 | |
| Gemma 4 31B | 79.46 | |
| GLM 5.2 FP8 | 76.5 | |
| Google Translate | 68.2 |
What is it?
North Small Translate is Cohere Labs' first model in the North family aimed squarely at machine translation, and the weights are public on Hugging Face. The model is a sparse mixture-of-experts decoder with 218B total parameters but only 25B active per token, so it runs far cheaper than its size suggests. It handles text in and text out across 50 languages, with a 16K-token limit on each side.
How does it work?
Sparse routing is what keeps the cost down: only 25B of the 218B parameters fire for any given token. Cohere ships three quantization builds — BF16, FP8, and NVFP4 W4A16 — so the smallest one fits on a single B200 or two H100s. An optional agentic mode runs a second pass over the output to find and fix its own errors, which lifts the WMT26 score from 83.60 to 84.36.
Why does it matter?
Translation buyers have had two realistic choices for years: a closed API from DeepL or Google, or a general-purpose LLM that treats translation as a side job. Cohere Labs now puts a model that outscores DeepL NextGen on WMT26 into researchers' hands under an open licence. The gap is widest on long documents, where North Small Translate scores 48.9 against Google Translate's 21.3.
Who is it for?
localisation teams, multilingual NLP researchers
Frequently asked questions
- Can I use North Small Translate commercially?
- North Small Translate ships under CC BY-NC 4.0, which covers research and non-commercial use only. Cohere routes commercial deployments through RWS Language Weaver, whose platform adds the security, scale and localisation tooling that enterprises need. The Hugging Face weights are also gated: downloaders have to agree to share contact details and accept Cohere Labs' Acceptable Use Policy.
- What hardware does North Small Translate need?
- Cohere publishes three builds of North Small Translate. BF16 needs four B200s or eight H100s, FP8 needs two B200s or four H100s, and the NVFP4 W4A16 build drops the floor to a single B200 or a pair of H100s. Sparse routing means only 25B of the 218B parameters are active per token, so throughput stays high.
- How does North Small Translate compare to DeepL and Google Translate?
- On the WMT26 all-languages score, North Small Translate reaches 83.60 against DeepL NextGen's 81.37 and Google Translate's 68.20. Its agentic mode, which reviews and repairs its own output, reaches 84.36. The lead grows on long documents: Cohere's long-context evaluation puts North Small Translate at 48.9, more than double Google Translate's 21.3.
- Can I try North Small Translate without downloading the weights?
- Yes. Cohere serves North Small Translate on its own API under the model ID north-small-translate-1-0, reachable through the Chat endpoint, so a short API call is enough to test it. Cohere's model documentation lists the model as live with a 16K context length, and points production users to Model Vault for dedicated deployment.
Try it
Load CohereLabs/North-Small-Translate-1.0 from Hugging Face, or call model north-small-translate-1-0 on the Cohere Chat API