Cohere's North Small Translate wins one of the five columns in its own results table
Cohere released North Small Translate on 10 September 2026: a 218B-parameter mixture-of-experts translation model with 25B active, open weights under CC BY-NC 4.0, built with RWS's Language Weaver team. Its headline, 83.6 on WMT26 across all languages, is the best number in its own paper. It is also the only column of five where it finishes first. On terminology it ranks 6th of 14, on long documents 2nd behind DeepL (48.9 vs 56.5), on structured JSON/HTML 4th, and on xCOMET-XL 5th, below Google Translate.
What Cohere released on 10 September 2026, and under what licence
10 September 2026. Cohere published North Small Translate, its first translation model in the North family, with weights on Hugging Face as CohereLabs/North-Small-Translate-1.0. The technical report, arXiv:2609.13916 (Kocmi et al., 23 authors), followed on 12 September. The model was submitted to the WMT 2026 General MT shared task under the name Cohere CAT+.
The facts that matter for procurement, from the Cohere changelog and the Hugging Face model card: 50+ languages (the paper lists 36 pretraining languages, though its text says 35, plus 14 added in post-training), 16K input and 16K output tokens, text only. Weights ship in W4A16 (two H100s or one B200), FP8 (four H100s or two B200s) and BF16 (eight H100s or four B200s). The licence is Creative Commons Attribution-NonCommercial 4.0, the repository is gated behind a contact-details form, and commercial use goes through Cohere's Model Vault, as Unite.AI reported the same day. The model is also on Cohere's free API tier as north-small-translate-1-0.
So a localization team can download it, benchmark it on its own content and publish the result. It cannot put it into a paid pipeline without a second contract. The New Stack ran that point as its headline on 13 September.
Originally, the Transformer paper was proposed to improve Google Translate. 9 years later, as the AI industry is built on top of its architecture, we've never stopped thinking about translation. Meet North Small Translate: a leading open machine translation model.
Cohere (@cohere) 10 September 2026
Cohere's own table: first on one column, second to sixth on the other four
Table 1 of the report compares 16 systems under 1T parameters across five evaluations. Two rows (Nemotron 3 Ultra and Qwen 3.6 27B) print three numbers for four remaining columns, so we cannot say which cell is empty; we exclude them from the ranks below and reproduce them as printed.
| System | Params total (active) | GEMBA WMT26 | xCOMET-XL WMT24++ | Terminology | Long context | Structured | Source |
|---|---|---|---|---|---|---|---|
| North Small Translate (Agentic) | 218B (25B) | 84.4 | 80.7 | 87.9 | 47.8 | 93.3 | Table 1 |
| North Small Translate | 218B (25B) | 83.6 | 80.6 | 89.7 | 48.9 | 93.7 | Table 1 |
| Mistral Large 3 | 675B (41B) | 81.6 | 78.3 | 91.2 | 20 | 93.8 | Table 1 |
| Qwen 3.5 397B | 397B (17B) | 81.6 | 78.5 | 87 | 19.5 | 59.4 | Table 1 |
| DeepL NextGen | API | 81.4 | 81.8 | 47.8 | 56.5 | 64.5 | Table 1 |
| Gemma 4 31B | 31B (31B) | 79.5 | 79.2 | 92.1 | 19.4 | 95.1 | Table 1 |
| Inkling Small | 276B (12B) | 78.5 | 78.7 | 91.1 | 21.5 | 91.3 | Table 1 |
| Muse Glimmer 30B | 30B (30B) | 78.2 | 77.7 | 90.1 | 32.5 | 91.7 | Table 1 |
| Command A+ | 218B (25B) | 76.5 | 81 | 68.7 | 30.5 | 73.7 | Table 1 |
| GLM 5.2 | 744B (40B) | 76.5 | 78.8 | 78.2 | 28.6 | 92.1 | Table 1 |
| Mistral Medium 3.5 | 128B (128B) | 75.7 | 76.7 | 90.3 | 19.3 | 94.2 | Table 1 |
| GPT-OSS 120B | 117B (5B) | 72.3 | 76.6 | 83.8 | 20.8 | 93.2 | Table 1 |
| Command A Translate (2025) | 111B (111B) | 70 | 76.9 | 56.9 | 0 | 79.5 | Table 1 |
| Google Translate | API | 68.2 | 80.8 | 38.1 | 21.3 | 48.8 | Table 1 |
| Nemotron 3 Ultra 550B | 550B (55B) | 67.1 | three values printed for four columns (84.8, 19.7); not assignable from the table as published | Table 1 | |||
| Qwen 3.6 27B | 27B (27B) | 61.9 | three values printed for four columns (89.2, 28.9); not assignable from the table as published | Table 1 | |||
Four readings follow directly from the numbers.
The agentic mode costs glossary adherence. The three-pass "translate, find errors, post-edit" variant gains 0.8 on GEMBA (84.4 vs 83.6) and loses 1.8 on terminology (87.9 vs 89.7), 1.1 on long context and 0.4 on structured output. A post-editing pass that optimises general quality is rewriting prescribed terms. That matches what the Barcelona Supercomputing Center found at WMT26 with CometKiwi, where the quality metric preferred the edit that broke the glossary (our 17 September note).
DeepL wins long documents and loses glossaries. DeepL NextGen posts the best long-context score in the table, 56.5, and a terminology score of 47.8, below everything except Google Translate (38.1). Cohere's footnote says DeepL is accessed by API, so "we cannot confirm implementation: for instance, whether it might pre-parse the source text". Cohere's launch blog says North "scores 48.9 on our long-context evaluation ... ahead of every general-purpose LLM we tested". That sentence is true and leaves out the one system in the table that beat it.
xCOMET-XL barely separates anyone on WMT24++. All 14 complete rows sit between 76.6 and 81.8, a 5.2-point band, and Google Translate at 80.8 sits above North at 80.6 while trailing it by 15.4 points on GEMBA. Cohere's explanation is saturation, and its Table 5 backs it with a human ESA study: 15.8% of WMT24 segments carried at least one major error from the same system, against 24.0% on WMT26. Korean runs the other way (29.9% vs 19.3%), so the saturation is not uniform.
Command A Translate, Cohere's 2025 model, scores 0.0 on long context. The new model's 48.9 is a real fix for a real regression, and it is Cohere's own previous release that makes the gap visible.
The model was trained toward the kind of judge that ranks it first
The WMT26 number is not a human ranking. Cohere scores every system with a reference-free GPT-5.6 Sol judge asked to rate adequacy and fluency separately, then averages the two. The official WMT26 human evaluation (Kocmi et al., 2026, "Contrastive dynamic human evaluation at scale") is cited as forthcoming. The report discloses, in its first footnote, that its main author is also an organiser of the WMT General MT task.
Cohere justifies the judge with a meta-evaluation on 24,000 internally human-labelled segments across 10 languages: pairwise accuracy with ties is 51.0 for its fluency-plus-adequacy GEMBA-ESA prompt, 50.6 for plain GEMBA-ESA, 44.9 for xCOMET-XL and 43.7 for spBLEU. That is a real argument that the LLM judge tracks humans better than xCOMET does. It is also a 6.1-point accuracy edge, on data that is not public, for a metric family the model was filtered, distilled and rewarded against at four earlier stages.
The report is candid about the one place the judge demonstrably failed. The online RL stage with an LLM-as-judge reward took the 14-language WMT25 average from 80.1% to 80.0% and "degraded the performance over few low-resource languages, which is due to weakness of the LLM judge on these languages". Cohere repaired it with a final DPO step at reduced learning rate and 5 warm-up steps, reaching 81.8%. A judge that is weak on a language cannot certify the model on that language either.
The blog and the paper do not quite agree on these. The blog gives Europe as 82.2 against Gemma's 73.9, and in successive paragraphs gives Gemma's South Asia score as 86.7 and as 88.04. We could not reconcile the groupings. The paper's Table 4 is the version we reproduce.
If you buy or build machine translation, pick the column before the model
Glossary-heavy content (product UI, regulated documentation). The table points to Gemma 4 31B (92.1 terminology, 95.1 structured) or Mistral Large 3 (91.2, 93.8) before North (89.7, 93.7), and away from DeepL (47.8) and Google Translate (38.1) on this particular benchmark. Do not turn on an agentic post-edit pass for this content without re-measuring term success; Cohere's own numbers show it costs 1.8 points.
Long documents (contracts, manuals, subtitles as one file). DeepL leads at 56.5 and North is the best open model at 48.9. Apart from North's own agentic variant (47.8), every other complete row is at 32.5 or below, and Gemma 4 31B, the terminology leader, scores 19.4. The best glossary model is one of the worst long-document models.
General content across many locales. This is where North's 83.6 applies, and where the cost claim sits: Cohere's blog puts a commercial-licence configuration at "80.1 score at just $0.000676 per task, using only 661 tokens on average", against $0.038928 per task for Gemini 3.1 Pro Preview. Which configuration scores 80.1, when the released model scores 83.6, is not reported.
Licensing is a gate, not a detail. CC BY-NC means an agency, LSP or SaaS vendor evaluating it for billable work needs Model Vault terms first. Evaluation and research use is clear. Everything past that needs Cohere's commercial agreement in writing.
A judgement, marked as one: the most useful thing in this report is not the model. It is a single table in which one vendor shows its model losing four of five columns, which is the honest shape of MT evaluation in 2026. There is no overall winner. There is a winner per content type.
Check it yourself
# 1. Licence and gating, straight from the Hugging Face API
curl -s https://huggingface.co/api/models/CohereLabs/North-Small-Translate-1.0 \
| python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["gated"], [t for t in d["tags"] if t.startswith("license")], d["lastModified"])'
# printed on 19 Sep 2026: auto ['license:cc-by-nc-4.0'] 2026-09-10T15:29:25.000Z
# 2. The report, Table 1 on page 2, Tables 2-5 in Sections 2-3
curl -L -o north.pdf https://arxiv.org/pdf/2609.13916
python3 -c "import pypdf; print(pypdf.PdfReader('north.pdf').pages[1].extract_text()[:1800])"
# 3. Rank North in each column (rows copied from Table 1, complete rows only)
python3 - <<'EOF'
rows = {"North":(83.6,80.6,89.7,48.9,93.7),"North Agentic":(84.4,80.7,87.9,47.8,93.3),
"Mistral Large 3":(81.6,78.3,91.2,20.0,93.8),"Qwen 3.5":(81.6,78.5,87.0,19.5,59.4),
"DeepL":(81.4,81.8,47.8,56.5,64.5),"Gemma 4 31B":(79.5,79.2,92.1,19.4,95.1),
"Inkling Small":(78.5,78.7,91.1,21.5,91.3),"Muse Glimmer":(78.2,77.7,90.1,32.5,91.7),
"Command A+":(76.5,81.0,68.7,30.5,73.7),"GLM 5.2":(76.5,78.8,78.2,28.6,92.1),
"Mistral Medium 3.5":(75.7,76.7,90.3,19.3,94.2),"GPT-OSS 120B":(72.3,76.6,83.8,20.8,93.2),
"Command A Translate":(70.0,76.9,56.9,0.0,79.5),"Google Translate":(68.2,80.8,38.1,21.3,48.8)}
for c, name in enumerate(["GEMBA","xCOMET","Terminology","Long ctx","Structured"]):
print(name, 1 + sum(v[c] > rows["North"][c] for v in rows.values()))
EOF
# GEMBA 2 / xCOMET 5 / Terminology 6 / Long ctx 2 / Structured 4To test the model on your own content, the free API tier serves it as north-small-translate-1-0. The GEMBA-ESA prompt Cohere used is printed in full in the report's Appendix A, so you can score your own outputs with the same judge family and compare against a human review of the same segments. That comparison, not the leaderboard, is the number that should decide a deployment.
What would prove this wrong
The claim is that North Small Translate's WMT26 lead depends on the choice of judge and task, and that no single score ranks it first for localization work. It is wrong if the official WMT26 General MT human evaluation (Kocmi et al., to appear in the Proceedings of the Eleventh Conference on Machine Translation) places Cohere CAT+ alone in the top significance cluster for at least 80% of the language pairs it entered. We will check this against the findings paper when it is published, and no later than 31 December 2026.
A second, narrower prediction: an independent evaluation of long-document translation published before 30 June 2027 will again place DeepL above North Small Translate. If a third party measures North ahead of DeepL on documents of 5,000 words or more, the long-context recommendation above is wrong.
Sources
- Tom Kocmi et al. (Cohere), North Small Translate: Advanced Cost-Effective Translation (Cohere CAT+), arXiv:2609.13916v1, 12 September 2026. Table 1 (five-column comparison), Table 2 (training progression), Table 3 (metric meta-evaluation, 24k segments), Table 4 (language groups), Table 5 (test set saturation), footnotes 1 and 4.
- Cohere, Introducing North Small Translate, 10 September 2026. Licence, hardware, regional figures, throughput, cost per task, RWS partnership.
- Cohere, Announcing Cohere's North Small Translate, changelog. Quantisations and hardware per quantisation.
- Hugging Face, CohereLabs/North-Small-Translate-1.0, last modified 10 September 2026, licence cc-by-nc-4.0, gated.
- Jonas Reeve, Unite.AI, Cohere Debuts Open-Weight 218B Mixture-of-Experts Machine Translation Model, 10 September 2026. Model Vault commercial licensing, API model id, 128 experts with 8 active.
- Adrian Bridgwater, The New Stack, Cohere's new translation model is open weights, but not for commercial use, 13 September 2026.
- Cohere (@cohere), launch post on X, 10 September 2026.
Related BLOMEGA guides: WMT 2026 terminology and the hard-example filter · LLM judges rank machines above human translators · Multilingual LLM judges and translationese bias
FAQ
What is Cohere North Small Translate?
An open-weight machine translation model Cohere released on 10 September 2026: a mixture-of-experts transformer with 218B total and 25B active parameters, covering 50+ languages, 16K input and output tokens, built with RWS's Language Weaver team. It scores 83.6 on WMT26 under a GPT-5.6 Sol judge, the top score in Cohere's own comparison.
Can I use North Small Translate commercially?
Not under the open licence. The Hugging Face weights are CC BY-NC 4.0 and gated. Commercial use requires a separate licence through Cohere's Model Vault. The model is also available on Cohere's free API tier as north-small-translate-1-0 for evaluation.
Is North Small Translate better than DeepL?
It depends on the task. In Cohere's own Table 1, North leads DeepL NextGen on WMT26 GEMBA (83.6 vs 81.4) and on terminology (89.7 vs 47.8), but DeepL leads on long-document translation (56.5 vs 48.9) and on xCOMET-XL WMT24++ (81.8 vs 80.6).
Which open model is best for glossary and terminology adherence?
In the same table, Gemma 4 31B scores highest on terminology (92.1) and on structured JSON/HTML translation (95.1). North Small Translate is sixth of 14 on terminology at 89.7, and its agentic post-editing mode drops that to 87.9.
How much GPU does North Small Translate need?
Per Cohere's changelog: two H100s or one B200 at W4A16, four H100s or two B200s at FP8, eight H100s or four B200s at BF16.