Cohere's translation model beats DeepL. You can't sell it.

Cohere released North-Small-Translate-1.0 on September 10, and it is now the best-scoring open-weight machine translation model we know of. It beats DeepL NextGen. It beats Google Translate by fifteen points. It also ships under a non-commercial license, which means the version you can download is the version you cannot put in front of a paying customer. Both halves of that sentence matter if you sell across a border.

What actually happened

Per Cohere's announcement, North Small Translate is a sparse mixture-of-experts model: 218B total parameters, 25B active, 50 languages, 16k tokens in and 16k out.

The WMT26 All Languages scores Cohere published:

• North Small Translate: 83.60
• North Small Translate, agentic multi-pass: 84.36
• Qwen 3.5 397B: 81.56
• DeepL NextGen: 81.37
• Google Translate: 68.20

The agentic variant re-reads its own output and fixes what it finds — a second pass, not a bigger model. The regional split is more interesting than the headline: Cohere reports roughly 8 to 10 points over DeepL in South Asia and MENA, 4 to 5 in Southeast Asia, and 1 to 3 in East Asia. The gap is smallest exactly where the incumbents have spent the most money.

Two numbers we would hold onto. Throughput: 112 tokens per second against 81 for a comparable model at low concurrency. Long documents: 48.9 against Google Translate's 21.3 — the failure mode where a translator loses the thread halfway down a product page.

The model card is explicit: CC BY-NC 4.0, plus an acceptable-use addendum. Research and non-commercial use only. Commercial access runs through Cohere sales or the RWS Language Weaver platform. Self-hosting needs a B200, or two H100s at W4A4 quantization.

Why an open translation model matters for your business

If you run a storefront in more than one language, translation is already a recurring bill, and it is usually a per-character contract with a vendor who just got beaten on a public benchmark. That is leverage at renewal even if you never download a weight.

What it is not is a free swap. "Open weights" bought you the right to evaluate, not the right to ship. The honest version of this release is: you can now prove your incumbent is beatable, then negotiate. Run your own corpus — your SKUs, your support macros, your returns policy — through the open checkpoint and the agentic variant, measure it against what you pay for today, and take the result into the room.

And price the hardware floor. A B200 to serve your own translation is a real line item, and for most catalogs an API still wins. The model that changes your economics is not always the model you host.

Sources: Cohere: North Small Translate, Hugging Face model card.