What Cohere released
Cohere released North Small Translate 1.0 on 10 September, a dedicated machine-translation model whose weights can be downloaded from Hugging Face. It is a sparse mixture-of-experts model with 218 billion parameters in total, of which about 25 billion are active for any given token, and it covers more than 50 languages, according to Cohere’s announcement.
The model card, published under Cohere Labs, describes a decoder-only transformer with 128 experts, eight of which are activated per token alongside shared experts. It accepts up to 16,000 tokens of input and returns up to 16,000 tokens of output, and it handles text only. Cohere says the smallest quantised configuration runs on a single Nvidia B200 or two H100s; the full-precision BF16 version needs eight H100s or four B200s.
Open weights, but not open for business
The licence matters more than usual here. North Small Translate is published under Creative Commons CC BY-NC 4.0. The weights are open in the sense that anyone can download, inspect, fine-tune and run them locally. The “NC” means non-commercial: research and personal use are permitted, while building the model into a paid product or a company’s own workflow is not. That makes it open-weight, not open source in the usual sense of the term, and Unite.ai notes that access on Hugging Face is gated behind a contact form.
Businesses have a separate route. Cohere’s blog says commercial deployment runs through RWS, whose Language Weaver research teams worked on the model and whose Language Weaver product carries it. The model card tells would-be commercial users to contact Cohere’s sales team. Researchers get the model for free; paying customers go through Cohere and RWS.
The benchmark numbers, and who produced them
Cohere’s headline claim is that the model beats dedicated translation services. In its own evaluation on WMT26 test data, North Small Translate scored 83.60 across all languages, rising to 84.36 in an agentic multi-pass mode in which the model looks for errors in its own translation and fixes them. In the same table Cohere puts Qwen 3.5 397B at 81.56, DeepL NextGen at 81.37, Gemma 4 31B at 79.46, GLM 5.2 at 76.50 and Google Translate at 68.20.
These are vendor-reported figures. WMT is the shared task the research community uses to compare translation systems, but the scoring in this table is Cohere’s own: the blog says outputs were judged by GPT-5.6-Sol, a language model, not by human raters and not by the official WMT26 rankings. Nobody outside Cohere has yet reproduced the comparison. The regional breakdown, as reported by Unite.ai, is also less uniform than the headline number: North Small Translate led Gemma 4 by 82.2 to 73.9 on European languages, but Gemma 4 scored marginally higher in South Asia, 86.7 to 86.2.
Cohere also reports a long-context test in which each model translates two chapters of a book in a single call, with each paragraph scored separately by the xCOMET-XL metric. North Small Translate scored 48.9, against 21.3 for Google Translate and 19.4 for Gemma 4 31B. That is the widest gap in the release, and the one most relevant to translating whole documents rather than single sentences.
Speed and cost
Cohere says the model delivers about 1.4 times the output throughput of Gemma 4 31B at low concurrency. It also puts the cost of an average translation task at $0.000676, against $0.038928 for Gemini 3.1 Pro. Like the quality scores, those are Cohere’s figures from Cohere’s test setup.
What to watch
The open weights are the part that can be checked. Independent runs on public WMT26 data, and the shared task’s own results, will show whether the lead over DeepL and Qwen holds up outside Cohere’s LLM-judged scoring. The long-context result is the one most worth testing, because it is where Cohere claims the largest margin.
