Cohere has released North Small Translate, its first translation model in the North family, available for research and non-commercial use under a CC BY-NC 4.0 license. This mixture-of-experts model achieves an 83.6 score on the WMT26 All Languages benchmark, outperforming proprietary models like DeepL NextGen and Google Translate.

  • North Small Translate scores 83.60 on WMT26, beating Qwen 3.5 (81.56), GLM 5.2 (76.50), and Gemma 4 31B (79.46).
  • The Agentic variant reaches 84.36 by finding and fixing translation errors.
  • It delivers up to 1.4x higher throughput than Gemma 4 31B, generating 112 vs. 81 tokens per second at low concurrency.
  • Long-context evaluation scores 48.9, more than double Google Translate's 21.3.
  • Commercial licensing via RWS costs $0.000676 per task, significantly less than Gemini 3.1 Pro Preview.

The model is designed for high-throughput generation and cost efficiency, with weights available on Hugging Face in near-lossless quantizations.