Cohere has released North Small Translate, an open-weight sparse Mixture-of-Experts (MoE) machine translation model developed with Cohere Labs and RWS. The model supports 50 languages and scores 83.6 on the WMT26 benchmark across all languages.

  • Architecture: 218B total parameters with 25B active per token, using a decoder-only sparse MoE Transformer with 128 experts.
  • Benchmarks: Scores 83.6 generally and 84.36 in an agentic multi-pass mode, outperforming DeepL NextGen (81.37) and Google Translate (68.20).
  • Performance: Achieves 112 output tokens per second at low concurrency and handles long documents effectively, scoring 48.9 on two book chapters.
  • Deployment: Available via Cohere's API for free until rate limits, with self-hosting options including BF16, FP8, and NVFP4 W4A16 checkpoints optimized for NVIDIA B200 and H100 hardware.

The model is positioned as a cost-effective and high-throughput alternative for organizations requiring accurate global communication capabilities.