Cohere Releases North Small Translate, a 218B MoE Open-Weight Translation Model Under CC BY-NC
Cohere and Cohere Labs released North Small Translate on September 10, 2026—a 218B-parameter MoE translation model with about 25B active parameters—under CC BY-NC 4.0 on Hugging Face, with vendor-reported WMT26 scores led by an 83.6 all-languages figure.
Cohere published North Small Translate on September 10, 2026, as the first translation model in its North family. Weights are on Hugging Face under CohereLabs/North-Small-Translate-1.0 (plus FP8 and NVFP4 W4A16 variants) for research and non-commercial use under CC BY-NC 4.0, with Cohere Labs’ Acceptable Use Policy still applying. Commercial deployment is directed to Cohere sales and to RWS Language Weaver, Cohere’s stated development partner.
The model is a decoder-only sparse mixture-of-experts transformer: about 218 billion total parameters with roughly 25 billion active per token, 128 experts with eight activated plus shared experts, and a 16K-token input and 16K-token output window. The Hugging Face card lists support for 50 languages and an attention stack that interleaves sliding-window layers (window 4096, RoPE) with global layers at a 3:1 ratio, following the pattern Cohere attributes to Command A.
Scores and serving
Cohere’s launch blog reports a vendor WMT26 all-languages score of 83.60 for the standard model and 84.36 for an agentic multi-pass workflow that finds and fixes translation errors. The same post places those figures ahead of listed proprietary and open-weight baselines (including DeepL NextGen, Google Translate, Gemma 4 31B, Qwen 3.5 397B A17B, and GLM 5.2) under an evaluation that uses GPT-5.6-Sol as a judge. Those numbers are vendor-reported; independent leaderboard confirmation should be treated as separate.
Hardware guidance on the card and blog snapshot ranges from multi-GPU BF16 footprints (examples include 4×B200 or 8×H100) down to tighter FP8 and W4A16 recipes (including 1×B200 or 2×H100 for W4A16). Cohere documents Transformers loading tips, a Hugging Face Space demo, and vLLM serving with the cohere_melody parsers for structural markers in the reply format.
Primary sources for this brief are Cohere’s September 10 North Small Translate blog post and the Hugging Face model card for CohereLabs/North-Small-Translate-1.0.
- #Opensource
- #Products
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.