Cohere Releases North Small Translate Open-Weight MoE Model

On September 10, 2026, Cohere released North Small Translate, an open-weight mixture-of-experts machine translation model. [1][3] The text-only model features 218 billion total parameters with 25 billion active parameters, and supports an input and output context limit of 16,000 tokens. [1][3] It supports over 50 languages, including tier-one languages such as Japanese, Korean, Modern Standard Arabic, and Ukrainian. [2] The weights are licensed for non-commercial and research use under CC BY-NC 4.0, while enterprise users can purchase a commercial license to deploy the model through Model Vault. [1][2] The minimum hardware requirement for deployment is a single B200 GPU or two H100 GPUs using NVFP4 W4A16 quantization. [1]

Cohere reported an all-languages WMT26 score of 83.60 for the model, evaluated using GPT-5.6-Sol as a judge. [1] An agentic multi-pass variant, which Cohere states can identify and correct translation errors, achieved a higher score of 84.36. [1][3] In these evaluations, the standard model outperformed competitors such as DeepL NextGen, Qwen 3.5 397B A17B, and Google Translate. [1] Regionally, Cohere reported that the model beat DeepL NextGen in all tested non-European regions, including an advantage of roughly 8 to 10 points in South Asia and the Middle East and North Africa. [1]

North Small Translate positions Cohere to aggressively compete in the enterprise localization market against both proprietary APIs like DeepL and Google Translate, and open-weight models like Qwen and Gemma. By offering an open-weight 25B active parameter model that fits on a single B200 or dual H100s, Cohere is targeting organizations that require strict data sovereignty for sensitive internal communications and safety manuals. The inclusion of an agentic variant suggests a shift toward automated self-correction in translation workflows, where models assess and repair their own outputs. What the sources don't address: How the pricing for the enterprise commercial license via Model Vault compares to the API costs of DeepL NextGen or Google Translate.