Cohere has launched North Small Translate, a decoder-only sparse Mixture-of-Experts (MoE) model with 218 billion parameters. The architecture activates approximately 11.5% of weights per token, resulting in 25 billion active parameters. The model supports 50 languages and achieved a score of 83.6 on the WMT26 benchmark.
The model is available in two versions, including an Agentic variant capable of self-correcting errors through a multi-pass workflow. In internal tests using GPT-5.6-Sol as a judge, the standard model scored 82.17 on EU languages, outperforming Gemma 4 31B, which scored 72.73. The model also demonstrated superior performance on long documents, scoring 48.9 compared to 21.3 for Google Translate.
Regarding performance, North Small Translate achieved up to 1.4 times higher throughput than Gemma 4 31B. At low concurrency, the model produced 112 tokens per second, while at high concurrency, it reached 39 tokens per second. Cohere developed the model in collaboration with RWS, utilizing their Language Weaver expertise to improve real-world translation quality.
Cohere offers the model through its Chat V2 API, where it is free to use until rate limits are reached. For commercial use, the model requires a license, though it is available for non-commercial self-hosting. The company reports a cost of $0.000676 per task for the model, which it claims is significantly lower than competing solutions like Gemini 3.1 Pro Preview.
