Editorial illustration for Cohere's 218B North Small Translate Model Scores 83.6 on WMT26
Cohere's 218B Model Hits 83.6 on Translation Benchmark
Cohere's 218B North Small Translate Model Scores 83.6 on WMT26
Cohere and Cohere Labs put out a new open-weight translation model this week called North Small Translate, and the numbers attached to it are the kind that invite scrutiny. It's a sparse Mixture-of-Experts model with 218 billion total parameters, though only 25 billion activate for any given token. It handles 50 languages, Albanian to Vietnamese, and on Cohere's own WMT26 evaluation it posts an average score of 83.6 across the board. Cohere claims that beats DeepL, Google Translate, and open rivals like GLM 5.2 and Mistral Large 3.
The release matters beyond the leaderboard. It's the first translation-specific model in Cohere's North family, arriving after Tiny Aya and Command A Translate, and it was built with help from RWS, the company behind Language Weaver, whose scientists and language specialists worked on real-world quality rather than just benchmark performance. Deployment is flexible too: free access through Cohere's API up to rate limits, self-hosting for non-commercial use, or a commercial license.
There's also a symbolic thread here. Machine translation was the original proving ground for the Transformer architecture itself, back in 2017. Cohere's return to that problem, nine years later, comes with its own framing of why it matters now.
Cohere has released North Small Translate, an open-weight machine translation model from Cohere and Cohere Labs. It is a sparse Mixture-of-Experts (MoE) model with 218B total and 25B active parameters. It covers 50 languages, from Albanian to Vietnamese.
Why this matters
Cohere grading its own model against DeepL and Google Translate, with GPT-5.6-Sol as judge, is worth flagging before anyone treats 83.6 as gospel. That's not a knock on North Small Translate specifically, it's just how benchmark math works when the vendor writes the test and picks the referee. We'd want to see the same 50-language set run by an outside party, or at least a rival lab's own numbers on the identical prompts, before ranking this above Mistral Large 3 or GLM 5.2 in any real sense.
That said, 218B total with 25B active is a real architectural choice, and free access on Cohere's API (rate limits aside) makes this easy to test yourself rather than take on faith. For teams doing localization at scale, cost per token matters as much as any leaderboard score. The agentic self-correction pass is the more interesting piece here, since multi-pass error fixing is where translation quality actually gets made or lost. Worth running your own language pairs before deciding what any of this means for your pipeline.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Anthropic: Five Campaigns Launched 200 Million AI Distillation Attacks
From Archives AutoFyn Agents Use Orchestrator and Verifier for Complex Tasks
Common Questions Answered
What is the architecture of Cohere's North Small Translate model and how many parameters does it use?
North Small Translate is a sparse Mixture-of-Experts (MoE) model with 218 billion total parameters, but only 25 billion parameters activate for any given token. This efficient architecture allows the model to handle complex translation tasks while maintaining computational efficiency compared to dense models of similar scale.
How many languages does North Small Translate support and what is its WMT26 benchmark score?
North Small Translate supports 50 languages, ranging from Albanian to Vietnamese, providing broad multilingual translation coverage. According to Cohere's evaluation on the WMT26 benchmark, the model achieved an average score of 83.6 across all supported language pairs.
Why should the WMT26 benchmark results be viewed with caution according to the article?
Cohere evaluated its own model against competitors like DeepL and Google Translate while using GPT-5.6-Sol as the judge, which raises concerns about potential bias since the vendor created both the test and selected the evaluation referee. The article recommends waiting for independent third-party evaluations or results from rival labs using identical prompts before definitively ranking North Small Translate above other models like Mistral Large 3 or GLM 5.2.
What makes the sparse Mixture-of-Experts design of North Small Translate significant for translation tasks?
The sparse MoE architecture allows North Small Translate to maintain a large total parameter count of 218 billion while only activating 25 billion parameters per token, resulting in improved efficiency and faster inference times. This design enables the model to achieve strong performance across 50 languages without requiring the computational resources of a fully dense model of equivalent size.
Further Reading
- A leading sovereign open-weight machine translation model - Cohere
- North Small Translate | Cohere - Cohere Docs
- Release Notes - Cohere Docs
- CohereLabs/North-Small-Translate-1.0-fp8 - Hugging Face
- CohereLabs/North-Small-Translate-1.0-w4a16 - Hugging Face
