Cohere released North Small Translate, an open-weight machine-translation model, on September 10, 2026, according to a post on the company's blog. The model uses a sparse mixture-of-experts (MoE) architecture with 218 billion total parameters, of which 25 billion are active for any given token, and covers more than 50 languages.
The weights are published on Hugging Face under a CC BY-NC 4.0 licence, which permits non-commercial self-hosting and research use. Commercial licences are available through RWS, which sells the model as part of its Language Weaver product. Cohere lists a minimum inference requirement of one Nvidia B200 GPU or two H100s at W4A4 quantisation; MarkTechPost reported requirements of four B200s or eight H100s at BF16 precision, falling to one B200 or two H100s at NVFP4 W4A16.
Benchmark results, according to Cohere
Cohere says North Small Translate scores 83.6 on its own WMT26 evaluation across all covered languages, with an agentic variant reaching 84.36. According to the company, the model outperforms DeepL NextGen (81.37), Qwen 3.5 397B A17B (81.56) and Google Translate (68.20) on the same evaluation. MarkTechPost noted that the scores are vendor-reported and were graded by an outside model, GPT-5.6-Sol. On European languages Cohere reports 82.2, against 73.9 for Gemma 4 31B; on South Asian languages it reports 86.2, marginally behind the 86.7 it lists for Gemma 4 31B.
On a long-document task spanning two book chapters, Cohere reports 48.9 for its model against 21.3 for Google Translate and 19.4 for Gemma 4 31B. The company also reports an output throughput of 112 tokens per second at low concurrency against 81 for Gemma 4 31B, and 39 against 30 at high concurrency, and puts the average cost of a translation task at $0.000676 across an average 661 tokens, against $0.038928 for Gemini 3.1 Pro Preview.
Cohere published architectural detail alongside the release: the model uses 128 experts with eight activated per token, which MarkTechPost calculated as roughly 11.5 percent of the weights active at any point; a 16,000-token context window for both input and output; and sliding-window attention over 4,096 tokens interleaved with global-attention layers at a ratio of three to one.
Why it matters
Open-weight translation models that claim commercial-grade accuracy remain uncommon, as most enterprise translation tools are closed and sold by subscription. Cohere frames the release as a sovereign alternative, aimed at organisations that want multilingual capability without depending on a single external provider. Non-commercial users can use the model through Cohere's API free up to rate limits, or download the weights directly. Because the headline figures come from Cohere's own WMT26 test suite rather than an independently administered leaderboard, no third-party verification of the 83.6 score has been published.