North Small Translate 1.0: Cohere's Quietest Big Model Yet

North Small Translate 1.0 is a 218-billion-parameter mixture-of-experts translation model from Cohere and Cohere Labs, and it arrived the way very few frontier-scale releases do: with a release note and nothing else. Cohere's changelog dates the announcement to September 9, 2026, but the three checkpoints had already been sitting on Hugging Face behind a gate since August 14 — three and a half weeks of weights with no blog post, no press coverage, and no launch thread. When this was written the main repository showed zero likes, 26 downloads in the last month, and a single unanswered community discussion. A 218B model with a vendor-reported WMT26 score of 83.60 that almost nobody has noticed is worth explaining, so here is what the documentation and the repository actually establish, separated from what they do not.

The numbers Cohere put its name to

Start with the specs, because they are unambiguous and they are the part of this release that is fully knowable. North Small Translate 1.0 is a decoder-only sparse mixture-of-experts model — the same shape Cohere used for Command A+ in May 2026 — with 128 experts, eight activated per token, plus shared experts that run on every token. The router takes a sigmoid activation over the expert logits and normalizes across the top-k selection. Attention alternates in a 3:1 ratio between sliding-window layers (window 4,096, RoPE) and global attention layers that carry no positional embeddings at all.

• Parameters — 218B total, 25B active per token

• Context — 16K input and 16K output

• Languages — English plus 49 others, 50 in total

• Tier-one languages — Modern Standard Arabic, German, French, Japanese, Korean, Russian, Ukrainian

• Checkpoints — BF16, FP8, and NVFP4 W4A16

• Minimum hardware — BF16: 4×B200 or 8×H100; FP8: 2×B200 or 4×H100; W4A16: 1×B200 or 2×H100

• License — CC BY-NC 4.0, with commercial use routed to Cohere's Model Vault

• Reported quality — WMT26 83.60, rising to 84.36 with an agentic multi-pass translation workflow

Two of those lines deserve more than a bullet. The context window is 16K output, which is unusual for a translation model and is the difference between translating a paragraph and translating a whole procedure document in one pass. And the agentic figure — 84.36 against 83.60 — is Cohere describing a workflow, not a model: multiple passes, presumably with an agent deciding what to re-translate. That is the same idea Cohere sold as "Deep Translation" with Command A Translate in August 2025, and it recurs here at larger scale.

What the repository shows that the release note does not

The release note is a marketing document; the repository is the engineering one, and a few details there matter more than the headline specs. Serving is supported through vLLM with cohere_melody>=0.9.0, and Cohere recommends greedy decoding explicitly "matching production deployment" — a translation model that tells you to turn sampling off is a translation model aimed at determinism, not at creative writing. Output is wrapped in <|START_TEXT|> and <|END_TEXT|> markers, and parsing expects Cohere's melody library. Those markers are not registered as special tokens, so setting skip_special_tokens=True will not strip them — a small trap that will produce visible tag litter in a naive integration.

There is also a chat template with a default system instruction that presents the model under the name "North Small Translate" and states it is built by Cohere. Per-conversation system messages are appended to that default rather than replacing it — there is a separate platform_instruction_override slot if you need to displace it. It is the kind of plumbing that tells you this is a post-trained chat model wearing a translation task, not a classic encoder-decoder.

Where it sits in Cohere's line

The lineage is the most useful context for reading this release, and it is documented. Command A Translate, announced August 28, 2025, was Cohere's first machine translation model — 111B parameters, a 16K window split 8K in and 8K out, 23 languages, and a CC-BY-NC research release. North Small Translate 1.0 more than doubles the language count to 50, triples the active parameter count to 25B, keeps the total window at 16K but makes the whole of it usable for output, and swaps a single dense-ish model for a sparse MoE with the same footprint Cohere already uses elsewhere in the North line.

The honest reading is that this is a specialist derivative of Cohere's current MoE core, post-trained for translation, rather than a new architecture. The model card says as much in one line — it was "post-trained specifically for translation quality" — and discloses nothing about the training corpus, the token count, or the data pipeline. Command A Translate's technical report described a difficulty-filtering data technique in detail; nothing equivalent has been published for this model.

What is genuinely not confirmed

This is where a quiet release has to be handled carefully, because the absence of noise cuts in both directions. The following are things nobody outside Cohere currently knows:

• What WMT26 83.60 measures. No metric is named on the model card or in the release note, and no WMT26 results page has been published against this model. "83.60" without a metric name is a number, not a result.

• Whether the score reproduces. No independent group has run this model. Every figure in this article is Cohere's own.

• What the agentic multi-pass workflow actually is. The 84.36 figure is described in a single clause. It is a system-level claim attached to a model card, and the system is not specified.

• What a commercial license costs. Cohere directs enterprises to sales and Model Vault. No price is published on the docs page.

• Whether a blog post is coming. Cohere has a blog. Command A Translate got a full one. This model got a release note.

None of that is an accusation. It is the difference between a model you can evaluate and a product you can buy, and right now North Small Translate 1.0 is squarely the former — plus one thing you can use for free.

The free tier is the part you can actually test today

Unlike a pure weights drop, this one has a live surface: North Small Translate 1.0 is on Cohere's free tier through the Chat V2 API, and Cohere's release note says it is free for both trial and production keys until rate limits are reached. That is an unusually permissive evaluation window — you do not need a paid contract to put real traffic through it. The weights, meanwhile, are the FP8 set on Hugging Face under CC BY-NC 4.0, so self-hosting for commercial purposes is off the table without a Model Vault agreement.

The practical shape of the decision, then: evaluate through the API at no cost, and pay only if you decide to ship it. That is a low-risk trial, and it is the sort of thing a routing layer is built for. If you are already calling translation through OrcaRouter — one key across a catalogue of more than 200 models, provider list price passed through at 0% markup — the evaluable question is not "do I migrate" but "does this beat what I am already calling on my own documents," which you can answer by running the same text through both. A vendor that cuts its price is a live change on our side the same day, because there is no reseller spread layered on top. And for a model with a single unreproduced benchmark and no third-party evaluation, failover is the mechanism that lets you try it without betting a production path on it: your traffic keeps flowing to the model that already works while the new one is graded on real inputs.

What would turn this from a release note into a release

Three things would move North Small Translate 1.0 from a documentation entry to a real competitor, and all three are observable from outside. First, a named metric for the WMT26 figure — an 83.60 with a metric attached can be compared, and one without cannot. Second, a first independent reproduction, whether that comes from the WMT26 evaluation campaign proper or from a third party running FLORES-style tests on the FP8 weights. Third, an actual launch post, which would signal that Cohere intends to sell this rather than merely to have published it.

Until then, the correct posture is the one the evidence supports: a genuinely interesting 218B MoE translation specialist with a free evaluation path, fifty languages and a document-length output window, whose quality claims rest entirely on its maker's word. Cohere has shipped four translation-capable models in thirteen months and only one of them got a blog post. Which of those two facts is the more informative is the question worth sitting with.