Overview
North Small Translate 1.0 is a machine-translation model from Cohere Labs, released on 10 September 2026 as the first entry in the North Small Translate line. It is a sparse Mixture-of-Experts decoder with 218 billion total parameters, of which 25 billion are active per token, and it handles text in and text out across 50 languages with a 16K-token limit on each side.
Cohere reports an all-languages score of 83.60 on WMT26, ahead of Qwen 3.5 397B A17B at 81.56, DeepL NextGen at 81.37, Gemma 4 31B at 79.46, GLM 5.2 FP8 at 76.50 and Google Translate at 68.20. An agentic mode, which reviews and repairs its own output, raises the score to 84.36. On Cohere's long-context translation evaluation the model scores 48.9 against Google Translate's 21.3 and Gemma 4 31B's 19.4.
The weights are gated on the Hugging Face Hub under CC BY-NC 4.0, covering research and non-commercial use. Cohere serves the model free on its own Chat V2 API under the model ID north-small-translate-1-0, offers production deployment through Model Vault, and points commercial translation and localisation customers to RWS Language Weaver. Cohere's documentation names Modern Standard Arabic, German, French, Japanese, Korean, Russian and Ukrainian as tier-one languages.
| Released | 2026-09-10 |
|---|---|
| License | CC BY-NC 4.0 |
| Weights | Open weights |
| Parameters | 218B total / 25B active (sparse MoE) |
| Context | 16K |
| Max output | 16K tokens |
| Architecture | Decoder-only sparse Mixture-of-Experts Transformer specialised for machine translation. Published in three checkpoints: BF16, FP8 W8A8 in compressed-tensors format, and NVFP4 W4A16. |
| Modalities | Text |
| Status | Generally available |
Benchmarks
North Small Translate 1.0 vs leading translation systems, as published by Cohere
| Benchmark | North Small Translate 1.0 | Qwen 3.5 397B A17B | DeepL NextGen | Gemma 4 31B | GLM 5.2 FP8 | Google Translate |
|---|---|---|---|---|---|---|
| WMT26 (all languages) | 83.6 | 81.56 | 81.37 | 79.46 | 76.5 | 68.2 |
| Long-context translation | 48.9 | — | — | 19.4 | — | 21.3 |
This model's scores
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Beats DeepL NextGen and Google Translate on Cohere's published WMT26 all-languages score
- Sparse routing activates only 25B of 218B parameters per token, keeping inference cost well below the total size
- Long documents are the widest margin: 48.9 on Cohere's long-context evaluation against Google Translate's 21.3
- Open weights on Hugging Face in BF16, FP8 and NVFP4 W4A16, so a quantized build fits one B200 or two H100s
- Free to call on the Cohere Chat V2 API for evaluation, with no download needed
Best for
- Localisation pipelines that need a self-hosted engine instead of a closed translation API
- Long-document translation where sentence-level engines lose track of context
- Research on multilingual and machine-translation methods with downloadable weights
- Sovereign or on-premise deployments that cannot send source text to a third-party service
- Enterprise translation and localisation via RWS Language Weaver under a commercial licence
How to access
| Provider | Model ID |
|---|---|
| Cohere API (Chat V2) ↗ | north-small-translate-1-0 |
| Hugging Face (open weights) ↗ | CohereLabs/North-Small-Translate-1.0 |
FAQ
What is North Small Translate 1.0?
North Small Translate 1.0 is a machine-translation model released by Cohere Labs on 10 September 2026. It is a sparse Mixture-of-Experts decoder with 218 billion total parameters and 25 billion active per token, covering 50 languages with a 16K-token input and output limit. The weights are published on Hugging Face under CC BY-NC 4.0.
Can North Small Translate 1.0 be used commercially?
Not under the published weights. CC BY-NC 4.0 covers research and non-commercial use only, and the Hugging Face repository is gated behind a contact-details agreement plus Cohere Labs' Acceptable Use Policy. Commercial deployments go through Cohere's Model Vault or through RWS Language Weaver, which offers the model on its translation and localisation platform.
What hardware does North Small Translate 1.0 need?
Cohere publishes three checkpoints. The BF16 build is roughly 437GB. The FP8 W8A8 build halves that to about 220GB and needs two B200s or four H100s. The NVFP4 W4A16 build brings the floor down to a single B200 or two H100s, which Cohere's documentation gives as the suggested self-hosting configuration.
How does North Small Translate 1.0 compare to DeepL and Google Translate?
On Cohere's published WMT26 all-languages score, North Small Translate 1.0 reaches 83.60 against DeepL NextGen's 81.37 and Google Translate's 68.20; its agentic mode reaches 84.36. The gap widens on long documents, where Cohere's long-context evaluation puts North Small Translate 1.0 at 48.9 versus 21.3 for Google Translate.
How do I call North Small Translate 1.0 without downloading it?
Cohere serves North Small Translate 1.0 on its own API through the Chat V2 endpoint, using the model ID north-small-translate-1-0. Cohere's documentation lists the model as available on the free tier for evaluation, with a 16K context length, and directs production traffic to Model Vault for dedicated deployment.