Cohere's new translation model lists Malay, but its licence bars commercial use

A translation model that handles Malay, Mandarin and Tamil well is not a niche product in Malaysia. It is the shape of almost every public counter, call centre and government form in the country. Cohere released one on 10 September 2026, with benchmark numbers that put it ahead of DeepL and Google Translate.[1]

North Small Translate is a mixture-of-experts model with 218 billion total parameters and 25 billion active, a 16,000-token input and output window, and support for more than fifty languages. Cohere calls it sovereign and open-weight, and says it was built with the language technology firm RWS.[1][2]

Cohere says North Small Translate is the first translation model in its North family, building on the Aya multilingual work and the earlier Command A Translate. It can be downloaded from Hugging Face in several near-lossless quantisations, with a demonstration space and implementation guides alongside it.[1]

The published language list includes Malay, Chinese, Tamil, Indonesian, Thai and Vietnamese, along with the major European languages. For a Malaysian reader that combination is the point. Most commercial translation stacks handle Malay and Tamil markedly worse than they handle French or German.[2][1]

On the WMT26 all-languages benchmark Cohere reports 83.60 for North Small Translate, against 81.56 for Qwen 3.5 397B A17B, 81.37 for DeepL NextGen, 79.46 for Gemma 4 31B and 68.20 for Google Translate. An agentic variant that finds and fixes its own errors scores 84.36.[1]

Southeast Asia is reported only as a regional average, where Cohere claims roughly four to five points over DeepL. No Malay-specific figure appears anywhere. Cohere also states in a footnote that the WMT scoring used a large language model as judge, which is not human evaluation.[1]

The Hugging Face release carries a CC BY-NC 4.0 licence. That permits research and non-commercial use and forbids commercial use, which rules out a Malaysian SME translating its product catalogue, a call centre routing customer messages, or an agency serving citizens with it under that licence.[1][2]

The commercial route is RWS and its Language Weaver product, the same partner whose research teams helped shape the model. No price, contract term or regional availability for that route appears in the post, and Cohere does not say whether Malaysian buyers are served today.[1]

Cohere lists a minimum of one NVIDIA B200, or two H100s, both at four-bit quantisation. That is data-centre hardware, not a workstation. The efficiency case is real once you clear that bar: 112 output tokens per second against 81 for Gemma 4 31B at low concurrency.[1]

On cost, Cohere puts an 80.1 score at 0.000676 US dollars per task using 661 tokens on average, against 0.038928 dollars for Gemini 3.1 Pro Preview. Those are the vendor's own figures, and they assume you already have somewhere to run the model.[1]