Translation AI doesn't usually make headlines, but Cohere's North Small Translate launch this week deserves a look — especially if you work with Indian languages or run any kind of multilingual content operation. Cohere, the Canadian AI company behind the Command and Aya model families, has shipped its first dedicated translation model, and the numbers it's claiming are bold enough to make Google and DeepL sit up.
01.What is North Small Translate?
It's a translation-only model built with RWS, a long-established language technology and localisation company. Under the hood it's a mixture-of-experts model with 218 billion total parameters and 25 billion active, with a 16,000-token context window for text in and out. That "small" in the name is relative — it's built to do one job very well rather than chat about everything.
Availability is split by use case: open weights for research and non-commercial use, while commercial deployment runs through Cohere's Model Vault and RWS's Language Weaver platform. That gives it a route to enterprise buyers who already use specialist localisation platforms rather than general-purpose chatbots.
02.The benchmark story
On Cohere'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, plus several open-weight models including Gemma 4 31B, GLM 5.2 and Qwen 3.5 397B. There's also an "agentic" version that can spot and correct its own translation errors — that one scored 84.36.
A caveat worth stating plainly: these are Cohere's own numbers, not independent results. But the margins, particularly the 15-point gap over Google Translate, are hard to ignore.
Two other numbers stood out to me. On long-context translation — full documents rather than sentences — it scored 48.9 against 21.3 for Google Translate. And on output throughput, it produced 112 tokens per second versus 81 for Gemma 4 31B in identical test conditions. Anyone who has watched a document get mangled sentence-by-sentence in an online translator knows why the long-context number matters.
03.Why the South Asia scores matter for India
Here's the part that made me actually pay attention. In South Asia, North Small Translate scored 86.2 — its strongest region alongside Europe — and Cohere says its biggest leads over DeepL came in South Asia and MENA, by roughly eight to ten points. For a country where millions of users think, text and transact in Hindi, Tamil, Bengali, Marathi, Telugu and dozens more languages, a serious competitor in this space is genuinely good news.
It also fits a pattern we've been tracking. India's own AI stack is being built multilingual-first — we recently wrote about BodhanAI's open models for 22 Indian languages, and earlier about Sarvam's Indus chat app for Indian-language users. Now a global player is treating South Asian language quality as a headline feature rather than an afterthought. That competition should improve translation quality across the board.
04.The price is the quiet headline
Cohere says 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) — against $0.038928 per task for Gemini 3.1 Pro Preview in its comparison. That's roughly 57x cheaper for comparable work. For startups and content teams translating lakhs of words a month, this is the difference between translation being a line item and being a rounding error.
Cost discipline is becoming the defining skill in AI work generally — the same principle applies whether you're choosing a translation model or running GPT-6 Astra agents. Our Zero to Hero guide has a full chapter on cost optimisation for exactly this reason.
05.Sovereign AI, open weights and what comes next
Cohere is pitching this as "sovereign multilingual AI" — giving customers control over where and how translation workloads run, which matters for governments and enterprises with data residency rules. That pitch lands well in India, where data localisation is an increasingly hot topic. The open weights (even research-only) also mean Indian universities and labs can study and build on the model rather than just consuming an API.
Translation is becoming the new battleground for specialised AI: instead of one giant model doing everything adequately, we're getting smaller, sharper models beating the giants at specific jobs. Expect more of this pattern across coding, legal, medical and other verticals.
Key Takeaways
- North Small Translate is Cohere's first dedicated translation model, built with RWS for 50+ languages
- MoE architecture: 218B total parameters, 25B active, 16K context
- Internal WMT26 score of 83.6 vs DeepL NextGen's 81.37 and Google Translate's 68.20
- Strongest regions: South Asia (86.2) and Europe — the agentic self-correcting version scores 84.36
- ~$0.0007 per task claimed cost, with open weights for research use
Frequently Asked Questions
What is Cohere North Small Translate?
It is Cohere's first dedicated translation model, built with language-tech firm RWS, supporting 50+ languages with a 218B-parameter mixture-of-experts design (25B active, 16K context).
How does it compare with DeepL and Google Translate?
It scored 83.6 on Cohere's internal WMT26 tests, ahead of DeepL NextGen (81.37) and Google Translate (68.20), and was strongest on South Asian languages at 86.2.
How much does it cost, and can I use it?
Cohere claims roughly $0.0007 per task, with open weights available for research use.
What This Means for You
If you run a business that touches multiple Indian languages — customer support, content, e-commerce listings, education material — put translation APIs on your evaluation list this quarter. The price-performance curve just bent in your favour, and even if you don't switch, the competition will force your current provider to respond. For students and freelancers, multilingual AI services (subtitling, localisation, document translation) just became dramatically cheaper to deliver. And if you want to build something around this, the skills are the same ones we teach in the GPT-6 Astra — Zero to Hero Guide — prompting, agents and cost control. Don't sleep on translation AI; it's less glamorous than humanoid robots, but it will quietly change how half of India uses the internet.