Cohere launches North Small Translate in translation push

Cohere has launched North Small Translate, its first translation model in the North family.

Developed with RWS, the model supports translation across more than 50 languages. It is available as an open-weight model for research and non-commercial use, while commercial deployment is offered through Cohere Model Vault and RWS's Language Weaver platform.

North Small Translate uses a mixture-of-experts architecture with 218 billion total parameters, including 25 billion active parameters. It is optimised for machine translation, supports text input and output, and has a context length of 16,000 tokens for both.

Cohere is entering a crowded translation market led by large technology groups and specialist providers. In the company's internal WMT26 benchmark testing, North Small Translate scored 83.6 across all languages, ahead of DeepL NextGen at 81.37 and Google Translate at 68.20, as well as several open-weight models including Gemma 4 31B, GLM 5.2 and Qwen 3.5 397B A17B.

An agentic version of the model, which Cohere said can identify and correct translation errors, scored 84.36. The company added that the standard model performed strongly across 32 high-resource languages and 18 additional languages, with less regional variation than rival systems of a similar size.

Benchmark claims

The company's figures suggest particular strength in Europe and South Asia. In Europe, North Small Translate scored 82.2 against 73.9 for Gemma 4 31B, while in South Asia it scored 86.2 against 86.7.

Cohere also said the model outperformed DeepL NextGen in the non-European regions covered by its testing, including the Middle East and North Africa, South Asia, South-East Asia and East Asia. It said the largest margins over DeepL were in South Asia and MENA, where the lead was roughly eight to 10 points.

The launch expands Cohere's multilingual product line, which already includes the Aya and Command families. Translation is becoming a more competitive area in generative AI as companies try to combine broad language support with lower running costs and greater control over deployment.

Speed and cost

Cohere said North Small Translate was designed for high output throughput. In tests on identical hardware and concurrency settings, the company said the model delivered up to 1.4 times the output throughput of Gemma 4 31B, reaching 112 output tokens per second at low concurrency versus 81 for Gemma, and 39 versus 30 at high concurrency.

That performance matters most for long documents, where machine translation systems often struggle to maintain quality over extended passages. On Cohere's long-context evaluation, North Small Translate scored 48.9, compared with 21.3 for Google Translate and 19.4 for Gemma 4 31B.

The company also positioned the model as a lower-cost option for commercial users. It said North Small Translate delivered a score of 80.1 at a cost of USD $0.000676 per task, based on average token use of 661, compared with USD $0.038928 per task for Gemini 3.1 Pro Preview.

Such claims are likely to appeal to organisations handling large multilingual workloads, especially those with data residency or deployment requirements that make open-weight models more attractive than closed API services. Cohere said the model supports what it described as more sovereign multilingual AI by giving customers greater control over where and how translation workloads run.

RWS role

RWS, which worked with Cohere on the model, is a long-established provider of language technology and translation services. Cohere said Language Weaver's research teams, science teams and language experts helped shape the system's real-world translation performance during development.

Commercial customers seeking managed deployment will be able to access the model through Language Weaver as well as Cohere's own model distribution channel. That gives Cohere a route to enterprise translation buyers that already use specialist localisation platforms rather than general-purpose AI services.

The launch also signals a broader push by model developers into more narrowly defined business tasks, where vendors believe specialist systems can outperform general-purpose chat models on quality, speed and operating cost. Translation is one of the more established categories, but it remains highly contested as providers seek to demonstrate gains through benchmark scores and lower prices.

North Small Translate is available now for research and non-commercial use, with commercial access offered through Cohere Model Vault and RWS's Language Weaver platform.