Cohere Releases Open-Weight Translation Model 'North Small Translate'… Surpasses Proprietary Models on WMT26

Cohere has released North Small Translate, an open-weight machine translation model utilizing a Mixture-of-Experts (MoE) architecture. The model supports over 50 languages and achieved an average score of 83.6 on the WMT26 benchmark, surpassing proprietary services such as DeepL and Google Translate, as well as major open-weight models like Gemma 4 and GLM 5.2.

Performance and Efficiency

The model secured higher scores than competing models based on GPT-5.6-Sol evaluation criteria. It demonstrated a significant performance gap compared to Gemma 4 31B, particularly in European languages, and scored 8–10 points higher than DeepL NextGen in the Middle East and South Asia regions. Its long-form translation capabilities are also outstanding, capable of translating the equivalent of two book chapters in a single call, achieving a score of 48.9 in this evaluation, which is the highest level among general LLMs.

Cost and Deployment

Designed for extreme cost-efficiency with enterprise deployment in mind. The average cost per task is $0.000676, significantly cheaper than Gemini 3.1 Pro Preview's cost per task of $0.038928. Throughput is up to 1.4x higher than Gemma 4 31B, with minimum hardware requirements of 1x B200 or 2x H100 GPUs (based on W4A4 quantization).

License and Partnership

Model weights are available for download on Hugging Face under the CC BY-NC 4.0 license, restricted to research and non-commercial use. Cohere co-developed the model with RWS, a translation specialist partnering with over 80% of the world's top brands, and enterprise security and scaling solutions will be provided through RWS's Language Weaver platform.