Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages Cohere released North Small Translate, an open-weight sparse Mixture-of-Experts translation model with 218B total and 25B active parameters covering 50 languages, which scores 83.6 on Cohere's WMT26 all-languages evaluation and 84.36 in an agentic multi-pass variant. Cohere says those results beat DeepL NextGen (81.37), Google Translate (68.20), GLM 5.2 FP8 (76.50) and Mistral Large 3, though the figures are vendor-reported with GPT-5.6-Sol as judge until independent WMT26 results appear. The model, built with RWS and the first translation model in Cohere's North family, is available free on Cohere's API within rate limits, for non-commercial self-hosting, or under a commercial license, with 16K input and 16K output token context. Cohere has released North Small Translate https://huggingface.co/CohereLabs/North-Small-Translate-1.0 , an open-weight machine translation model from Cohere and Cohere Labs. It is a sparse Mixture-of-Experts MoE model with 218B total and 25B active parameters. It covers 50 languages, from Albanian to Vietnamese. On Cohere’s WMT26 evaluation, it scores 83.6 averaged across all languages. Cohere says that beats DeepL and Google Translate, plus open options like GLM 5.2 and Mistral Large 3. Is it deployable? Yes. Call it free on Cohere’s API until rate limits, self-host it non-commercially, or license it commercially. Back to Where the Transformer Started Google researchers introduced the Transformer in 2017 with Attention Is All You Need https://arxiv.org/abs/1706.03762 . Its main results came from WMT 2014 English-to-German and English-to-French translation. 9 years later, Cohere is returning to that original problem with a dedicated model. Cohere’s launch post on X https://x.com/cohere/status/2098081558087270736 frames translation as a sovereignty issue. Organizations that cannot communicate globally cannot stay sovereign. North Small Translate is the first translation model in Cohere’s North family. It follows Tiny Aya https://cohere.com/blog/cohere-labs-tiny-aya and Command A Translate https://docs.cohere.com/docs/command-a-translate in Cohere’s multilingual lineage. Cohere built it with RWS https://cohere.com/blog/rws-and-cohere-build-ai-language-intelligence , whose Language Weaver scientists and language experts shaped its real-world quality. Architecture The model structure https://huggingface.co/CohereLabs/North-Small-Translate-1.0 describes a decoder-only sparse MoE Transformer. Here are the key details: - Experts: 128 experts, 8 activated per token, plus shared experts applied to every token. - Router: A sigmoid over expert logits, normalized over the selected top-k. - Attention: Sliding-window layers window 4096, RoPE and global layers without positional embeddings, interleaved 3:1. - Lineage: That attention layout was first introduced in Command A. - Context: 16K input and 16K output tokens, text only. - Training: Post-trained specifically for translation quality. About 11.5% of the weights are active per token. Per-token compute tracks the 25B active parameters. Memory still has to hold all 218B. Benchmarks Cohere team reports these WMT26 all-languages scores in its launch blog https://cohere.com/blog/north-small-translate : | Model | WMT26 score | |---|---| | North Small Translate Agentic | 84.36 | | North Small Translate | 83.60 | | Qwen 3.5 397B A17B | 81.56 | | DeepL NextGen | 81.37 | | Gemma 4 31B on | 79.46 | | GLM 5.2 FP8 | 76.50 | | Google Translate | 68.20 | The Agentic variant runs a multi-pass workflow that finds and fixes its own errors. Cohere’s scoring bands treat 80 to 100 as perfect or minor errors only. One caveat matters here. These are Cohere’s own runs, with GPT-5.6-Sol as the judge. Treat them as vendor-reported until independent WMT26 https://www2.statmt.org/wmt26/translation-task.html results appear. Regionally, both versions beat Gemma 4 31B on across Europe. On EU languages, the standard model scores 82.17 against Gemma’s 72.73. South Asia is close, at 86.16 for North against 88.04 for Gemma. Speed, Long Documents and Cost In Cohere’s tests, the model produced 112 output tokens per second against 81 for Gemma 4 31B. That was at low concurrency on identical hardware. At high concurrency, the figures were 39 against 30. Cohere calls this up to 1.4x higher throughput. Long documents are a stronger point. The model scores 48.9 when translating 2 book chapters in 1 call. Google Translate scores 21.3 and Gemma 4 31B scores 19.4. Quality is measured per paragraph with xCOMET-XL. In Cohere’s cost chart, the model scores 80.1 at $0.000676 per task, averaging 661 tokens. Gemini 3.1 Pro Preview high costs $0.038928 per task, about 58x more. Qwen 3.5 397B A17B costs $0.004525 and Command A+ costs $0.005158. How to Run It The fastest path is Cohere’s Chat V2 API https://docs.cohere.com/docs/north-small-translate-1.0 . The model is free there until rate limits: python from cohere import ClientV2 co = ClientV2 api key="