Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages

Cohere has put out North Small Translate, an open-weight machine translation model built by Cohere and Cohere Labs. It’s a sparse Mixture-of-Experts system with 218B total parameters and 25B active at a time, aimed squarely at one job: translation across 50 languages, from Albanian to Vietnamese.

The headline number is Cohere’s own WMT26 result: 83.6 averaged across all languages. The company says that tops DeepL, Google Translate, and several open models including GLM 5.2 and Mistral Large 3. There’s a catch, of course. These are Cohere-reported scores, judged with GPT-5.6-Sol, so they’re a vendor benchmark until someone else runs the same gauntlet.

Still, the model does more than win a blog post. Cohere says it can be used free through its API until rate limits kick in. It can also be self-hosted for non-commercial use, or licensed commercially. For deployment, that matters more than a flashy chart. Translation is useful when it is boring, dependable, and easy to plug into a workflow.

The model also comes with some practical muscle. In Cohere’s tests, it produced 112 output tokens per second on one setup versus 81 for Gemma 4 31B, and 39 versus 30 at higher concurrency. On long documents, it scored 48.9 on a two-chapter translation task, ahead of Google Translate at 21.3 and Gemma 4 31B at 19.4. Cohere also says the model reaches 80.1 at a cost of $0.000676 per task, far below Gemini 3.1 Pro Preview (high), which it puts at $0.038928.

Under the hood, this is a decoder-only sparse MoE Transformer with 128 experts, 8 active per token, plus shared experts. It uses a mix of sliding-window and global attention, and Cohere says the attention layout first appeared in Command A. The model has 16K input and 16K output tokens of text-only context, and Cohere is also offering three checkpoints for self-hosting, including a 4-bit version that runs on 1x B200 or 2x H100.