Cohere releases North Small Translate, an open-weight machine-translation model

Cohere released North Small Translate, a mixture-of-experts machine-translation model with 218 billion total parameters and 25 billion active, covering 50+ languages under a non-commercial open-weight license.

Cohere released a downloadable translation model for 50+ languages. The license is free for research, not for selling. The contest scores are Cohere’s, not settled science.

Cohere published “Introducing North Small Translate: A leading sovereign open-weight machine translation model.” That company blog is the filing event.

The model on the page is North-Small-Translate-1.0. It uses a mixture-of-experts architecture — many specialist sub-networks, only some of which fire on each token — with 218B total / 25B active parameters. That is 218 billion weights in the whole network, and 25 billion in use on a given token. Context is 16k input and 16k output (16,000 tokens each), text-in and text-out. Cohere says it supports 50+ languages; the full list is on Cohere’s page, not reprinted here.

The license is open weights for research and non-commercial use under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial). Open-weight means you can download the model file. Non-commercial means you cannot sell it under this license. Hugging Face weights — including near-lossless quantizations, which shrink the file with little quality loss — and a Hugging Face Space demo are linked from the post. This is not unrestricted commercial open source. Cohere points enterprises that need a commercial path to RWS Language Weaver.

Cohere’s WMT26 “All Languages” score claim — WMT is the Workshop on Machine Translation, a standard translation contest — is 83.6 for North Small Translate versus listed comparators on the page: DeepL NextGen 81.37, Google Translate 68.20, Gemma 4 31B (on) 79.46, GLM 5.2 FP8 76.50, Qwen 3.5 397B A17B 81.56. An “Agentic” variant that finds and fixes translation errors is listed at 84.36. Those scores are Cohere’s published evaluations; the page footnotes GPT-5.6-Sol as judge.

Company internal tests claim up to 1.4× higher output throughput than Gemma 4 31B TP1 under identical concurrency and hardware in Cohere’s chart — 112 vs 81 TOPS (tera-operations per second) at low concurrency; 39 vs 30 at high. 1.4× means about forty percent more output in the same time, on Cohere’s chart. Attribute as Cohere’s test, not an independent throughput audit.

The same company table claims 48.9 on Cohere’s book-chapter eval versus Google Translate 21.3 and Gemma 4 31B 19.4. That is Cohere’s long-context framing — not an external audit.

The setup details are in Sources. Hardware pins and bit-width cutoffs on the page are Cohere’s guidance, not a desk hardware check.

Cohere says the model was developed with RWS / Language Weaver research, science, and language-expert teams. Commercial enterprises are pointed to Language Weaver for secured deployment beyond the non-commercial Hugging Face weights. This filing does not name customer logos or a deal size.