Cohere Labs just released an open weight, 218 billion parameter translation model with a built in multi pass workflow that can meaningfully change output quality depending on how you call it. Here is what to test before you decide whether it replaces your current translation vendor.
Cohere Labs released North-Small-Translate-1.0 this week, a sparse mixture of experts translation model with 218 billion total parameters and 25 billion active parameters, supporting 50 languages, and it is worth a hands on look before your team writes off open weight translation models as good enough for casual use but not production quality.
The model supports a 16,000 token input and output context window, confirmed on Cohere's own release notes, which is large enough to handle full contracts, technical manuals, or long form articles in a single pass without the chunking logic most translation pipelines still carry over from older, smaller context models. According to Cohere's own architecture documentation, the model uses a decoder only design with 128 total experts, of which 8 activate per token, which is how a model this large keeps inference cost closer to what you would expect from a much smaller dense model.
The detail worth testing yourself is the agentic multi pass workflow. Cohere reported that a standard single pass run scored 83.60 on its internal WMT26 evaluation, rising to 84.36 when the multi pass mode is enabled, ahead of the 81.37 the company recorded for DeepL and 68.20 for Google Translate in that same vendor run comparison. Because those are Cohere's own benchmark numbers rather than an independent third party evaluation, treat the specific scores as a starting point rather than a verdict, and run your own comparison against whatever your team currently uses.
Practically, that means budgeting differently depending on how you deploy it. Single pass translation will be the cheaper, faster default for high volume, lower stakes content. The multi pass mode costs more, since it runs multiple forward passes per document, but is worth reserving for the language pairs or document types where your own testing shows single pass output falling short, rather than turning it on universally.
Steps to try this yourself before committing to anything in production:
1. Pull North-Small-Translate-1.0 from Hugging Face under the CohereLabs organization page for initial non commercial evaluation.
2. Feed it a real document from your own pipeline, not a benchmark sentence, and check whether the full document fits inside the 16,000 token window before you need chunking logic.
3. Run the same document through single pass mode first, and compare it against your current vendor's output on terminology consistency, not just fluency.
4. If quality is close but not quite there, re run the same document with the agentic multi pass workflow enabled and compare the cost delta against the quality gain for your specific content type.
5. Before moving anything to production, contact Cohere's sales team for a commercial license, since the public release is CC BY-NC 4.0 and covers non commercial use only.
Try it today
- Pull North-Small-Translate-1.0 from Hugging Face under the CohereLabs organization page for initial non commercial evaluation.
- Feed it a real document from your own pipeline and check whether it fits inside the 16,000 token context window before you need chunking logic.
- Run the same document through single pass mode first and compare it against your current vendor's output on terminology consistency.
- If quality is close but not quite there, re run with the agentic multi pass workflow enabled and weigh the added inference cost against the quality gain.
- Contact Cohere's sales team for a commercial license before production use, since the public release is CC BY-NC 4.0 and covers non commercial use only.