Cohere has released North Small Translate, a 218B parameter Mixture-of-Experts model that scores 83.6 on the WMT26 benchmark across 50 languages. The release claims this performance outperforms DeepL, Google Translate, GLM 5.2, and Mistral Large 3.
The model is available on Cohere’s API without cost until rate limits apply. Users may also self-host it for non-commercial use or purchase a commercial license.
Returning to the original problem
Google researchers published the Transformer architecture in 2017 using the “Attention Is All You Need” paper. Their primary results came from translating English to German and French on the WMT 2014 dataset. Nine years later, Cohere is addressing that same task with a dedicated system.
A post on X frames translation as a matter of sovereignty. The argument is that organisations cannot maintain sovereignty if they cannot communicate globally. North Small Translate is the first model in Cohere’s North family. It follows the Tiny Aya and Command A Translate models in the company’s multilingual line. Cohere built the system with RWS, whose Language Weaver scientists and experts shaped the real-world quality.
Technical details
The model structure is a decoder-only sparse Mixture-of-Experts Transformer. The key specifications include:
- Experts: 128 total experts, 8 activated per token, plus shared experts applied to every token.
- Router: A sigmoid function over expert logits, normalised over the selected top-k.
- Attention: Sliding-window layers with a window of 4096 and RoPE, interleaved with global layers without positional embeddings at a 3:1 ratio.
- Lineage: This attention layout was first introduced in Command A.
- Context: 16K input tokens and 16K output tokens, text only.
- Training: Post-trained specifically for translation quality.
Approximately 11.5% of the weights are active per token. Per-token compute tracks the 25B active parameters. Memory must still hold all 218B parameters.
Benchmark results
The Cohere team reported these WMT26 all-languages scores in its launch blog:
Model WMT26 score North Small Translate (Agentic) 84.36 North Small Translate 83.60 Qwen 3.5 397B A17B 81.56 DeepL NextGen 81.37 Gemma 4 31B (on) 79.46 GLM 5.2 FP8 76.50 Google Translate 68.20
The Agentic variant runs a multi-pass workflow that identifies and corrects its own errors. Cohere’s scoring bands treat 80 to 100 as perfect or minor errors only. One caveat applies here. These are Cohere’s own runs, with GPT-5.6-Sol acting as the judge. Treat them as vendor-reported until independent WMT26 results appear.
Regionally, both versions beat Gemma 4 31B (on) across Europe. On EU languages, the standard model scores 82.17 against Gemma’s 72.73. South Asia is close, at 86.16 for North against 88.04 for Gemma.
Speed, long documents and cost
In Cohere’s tests, the model produced 112 output tokens per second against 81 for Gemma 4 31B. That was at low concurrency on identical hardware. At high concurrency, the figures were 39 against 30. Cohere calls this up to 1.4x higher throughput.
Long documents are a stronger point. The model scores 48.9 when translating two book chapters in one call. Google Translate scores 21.3 and Gemma 4 31B scores 19.4. Quality is measured per paragraph with xCOMET-XL.
In Cohere’s cost chart, the model scores 80.1 at $0.000676 per task, averaging 661 tokens. Gemini 3.1 Pro Preview (high) costs $0.038928 per task, about 58x more. Qwen 3.5 397B A17B costs $0.004525 and Command A+ costs $0.005158.
How to run it
The fastest path is Cohere’s Chat V2 API. The model is free there until rate limits apply:
from cohere import ClientV2 co = ClientV2(api_key="") response = co.chat( model="north-small-translate-1-0", messages=[{"role": "user", "content": "Translate everything that follows into French:\n\nEnterprises need accurate translations of business-critical documents."}], ) print(response.message.content[0].text)
For self-hosting, Cohere publishes three checkpoints, the same ones it serves in production:
Checkpoint Blackwell Hopper BF16 4x B200 8x H100 FP8 2x B200 4x H100 NVFP4 W4A16 1x B200 2x H100
What it means
For people making things, this changes the cost of translation work. The 4-bit checkpoint runs on a single Blackwell B200 or two Hopper H100 cards. This allows teams to host the model locally without paying per-request fees. The speed gain of 112 tokens per second means long documents like contracts or technical manuals can be processed in one go. This removes the need to break text into small chunks for translation.
