Cohere North Small Translate Beats DeepL on WMT26

Cohere has launched North Small Translate 1.0, a text-only machine-translation model featuring downloadable weights under a Creative Commons Attribution-NonCommercial 4.0 license. Built as a sparse mixture-of-experts transformer, the model contains 218 billion total parameters but activates just 25 billion per token by routing inputs to eight of its 128 experts. To run the quantized NVFP4 W4A16 checkpoint on a single server, Cohere recommends utilizing either two Nvidia H100 GPUs or a single B200 GPU.

In evaluations on the WMT26 benchmark, North Small Translate achieved an all-language score of 83.60, surpassing DeepL NextGen at 81.37, Qwen 3.5 397B A17B by 2.04 points, and Google Translate at 68.20. An optional agentic multi-pass workflow, which detects and corrects its own translation errors at the cost of higher latency, raised this score to 84.36. In a separate long-context evaluation translating book chapters, the model scored 48.9 on the xCOMET-XL metric, significantly outperforming Google Translate at 21.3 and Gemma 4 31B at 19.4.

The model supports up to 16,000 input and output tokens, utilizing sliding-window attention layers with a 4,096-token local context window. Cohere trained the system using a five-stage protocol that combined supervised fine-tuning, direct preference optimization, and online reinforcement learning with RWS. To optimize training, developers employed difficulty sampling alongside language-specific supervision techniques, such as Best-per-Language Forward Translation and Post-Edit Driven Preference Distillation, to focus updates on complex documents.

For practitioners, the model offers a specialized alternative to massive, reasoning-heavy LLMs, allowing localization teams to control data placement and customize terminology prompts. While the model is free to use on the Cohere API up to certain rate limits under the identifier north-small-translate-1-0, commercial deployments require licensing through Cohere Model Vault or RWS Language Weaver. The single-pass setup suits high-volume workloads like catalog translation, while the agentic workflow can be reserved for high-value segments.