Summary: Cohere released North Small Translate on September 10, 2026 as a 218B-parameter MoE translator with 16k input and output contexts. Cohere reports a 83.60 WMT26 average, up to 1.4 times Gemma 4 31B TP1 output throughput, and much stronger long-context results, while noting that its scores are vendor-judged. Open weights are non-commercial, so production use requires the commercial path through RWS Language Weaver.
A translation pipeline rarely fails on a scoring gap; it fails on the long document. One that renders a paragraph beautifully can come apart when two chapters arrive in a single call, and the customary workaround — chunking — buys latency, cost, and a new family of boundary errors. North Small Translate, released by Cohere on September 10, 2026, sits on that seam.
What the spec actually says
A mixture-of-experts translator: 218B total parameters with 25B active, a 16k input context and a 16k output context, text in and text out. Cohere puts language coverage at 50+, ships several quantizations that come close to lossless, and offers the weights on Hugging Face under CC BY-NC 4.0 for research and non-commercial work. The hardware floor is 1× B200, or 2× H100s at W4A4.
Against WMT26, the all-languages figure Cohere reports is 83.60 — above Qwen 3.5 397B A17B (81.56); DeepL NextGen (81.37); Gemma 4 31B on (79.46); and Google Translate (68.20). The agentic variant, which hunts down and repairs translation errors, reaches 84.36. All come from Cohere itself, judged with GPT-5.6-Sol, so they mark where your own evaluation begins, not a final ranking.
Throughput, and what the long context buys
The claims a builder can act on are the serving numbers. With hardware and concurrency held equal, Cohere reports output throughput up to 1.4x that of Gemma 4 31B TP1: 112 against 81 output tokens per second under low concurrency, 39 against 30 under high concurrency. As a percentage that is a 30–38% edge in tokens per second, and it shows up most sharply once outputs grow long.
The long-context evaluation from Cohere translates two book chapters in one call and scores each paragraph with xComet-XL. North Small Translate lands there at 48.9 — over twice Google Translate (21.3), and more than twice Gemma 4 31B (19.4). No per-language or per-document-type split accompanies those figures in the supplied material, so what stays open is how documents of your own behave at that length.
Priced per task, not per token
Efficiency here is framed per task instead of per token. For a commercial license Cohere cites 80.1 on the score at $0.000676 per task, drawing on an average of 661 tokens; for Gemini 3.1 Pro Preview (high) the same per-task figure is $0.038928, with 5,762 tokens behind it. Qwen 3.5 397B A17B lands at $0.004525 per task, and Cohere’s own Command A+ at $0.005158. To compare vendors on equal footing, reproduce that measurement: push your document mix through every candidate and total what each finished translation costs, retries counted.
Where the license splits
Non-commercial is what the open weights permit. Enterprises needing a managed localization platform, security, and commercial rights go through RWS’s Language Weaver, a product built in partnership with RWS. That is the fork: Hugging Face is enough to start internal prototyping and research today, while anything headed for production needs a commercial agreement.
Models kept inside your own boundary and models that leave it carry their own tradeoffs, as zero data retention routing for AI APIs sets out. For a translation pipeline the choice is stark: keep the weights on your own infrastructure, or else accept a path the vendor hosts.
What to measure before you commit
Three checks repay the effort first: push your longest real documents through in one call; at the concurrency you really serve, measure throughput; and over your language mix, track cost per finished task. Per-language testing stops being optional here — in Cohere’s regional numbers, South Asia has the model merely level with Gemma 4 31B (86.2 against 86.7), while Europe has it ahead (82.2 against 73.9). Where quality lands on low-resource languages beyond the 50+ list goes unspecified in the supplied material, so verify coverage against the locales you actually ship.