Cohere is circling back to the Transformer’s roots with North Small Translate, a new open-weight model designed specifically for machine translation. Built as a sparse Mixture-of-Experts (MoE) architecture, it packs 218 billion total parameters but activates only 25 billion per token. This efficiency trick allows it to cover 50 languages, from Albanian to Vietnamese, while claiming an average score of 83.6 on Cohere’s WMT26 evaluation. The company asserts this performance surpasses established players like DeepL and Google Translate, alongside open-source rivals such as Mistral Large 3.
The real hook here isn’t just accuracy, but throughput and price. In internal tests running on identical hardware, North Small Translate generated 112 output tokens per second, edging out Gemma 4 31B’s 81 tokens. For enterprises drowning in long-form content, the model handles 16K context windows with notable stability, scoring 48.9 on two-book-chapter translations compared to Google Translate’s 21.3. Cost-wise, the difference is stark. Cohere lists the price at $0.000676 per task, which is roughly 58 times cheaper than Gemini 3.1 Pro Preview. You can access it freely via Cohere’s API until rate limits hit, or self-host the 4-bit checkpoint on a single B200 GPU.
However, seasoned engineers should pause before swapping out their current pipelines. The benchmark scores are vendor-reported and judged by GPT-5.6-Sol, not an independent third party. While the agentic variant boosts scores to 84.36 by fixing its own errors, the lack of external validation on the WMT26 suite leaves some room for skepticism. Still, framing translation as a “sovereignty issue” adds a layer of strategic weight to the release, suggesting this isn’t just about swapping words but maintaining organizational autonomy in global communications.
This move signals a shift away from generalist LLMs doing mediocre translation toward specialized, efficient models that respect budget constraints. By partnering with RWS and leveraging the Command A lineage, Cohere is betting that dedicated architecture beats brute force. Whether the community validates these speed claims remains to be seen, but the barrier to entry for high-quality, self-hosted translation just dropped significantly.
