Cohere North Small Translate: 218B MoE translation model

On 2026-09-10, Cohere published “Introducing North Small Translate” — its first translation model in the North family — and opened CohereLabs/North-Small-Translate-1.0 on Hugging Face under CC BY-NC 4.0. The model is a sparse Mixture-of-Experts transformer with 218B total parameters (25B active) across 50+ languages, and Cohere claims it beats DeepL NextGen and Google Translate on a self-run “WMT26 all-languages benchmark” using GPT-5.6-Sol as judge.

What happened

• Release date: 2026-09-10 (Cohere blog, 2026-09-10)
• License: CC BY-NC 4.0 with Acceptable Use Policy addendum (HF model card)
• Architecture: 218B total / 25B active sparse MoE — 128 experts, 8 activated per token, plus shared experts (HF model card)
• Context: 16K input, 16K output (HF model card)
• Languages: 50 across Europe, MENA, South Asia, Southeast Asia, and East Asia (full list in the model card)
• Partner: Built with RWS, specifically Language Weaver’s research and science teams (Cohere blog, 2026-06-01); commercial path is RWS Language Weaver or Cohere sales
• Lineage: First translation model in the North family; ancestors include Tiny Aya and Command A Translate (Cohere blog, 2026-09-10)

North Small Translate is gated behind a click-through on Hugging Face requiring agreement to the License Agreement, Cohere Labs AUP, and Cohere Privacy Policy.

Why it matters

Open-weight translation at sub-millidollar cost. Cohere reports a score of 83.60 on its own “WMT26 all-languages benchmark,” ahead of DeepL NextGen (81.37), Qwen 3.5 397B A17B (81.56), Gemma 4 31B (79.46), GLM 5.2 FP8 (76.50), and Google Translate (68.20) (Cohere blog, 2026-09-10). An agentic multi-pass variant scores 84.36. The commercial cost line is $0.000676 per task at an 80.1 score, compared to $0.038928 for Gemini 3.1 Pro Preview on the same run — a 5,762× difference per Cohere’s numbers.

Throughput claim. Cohere reports up to 1.4× higher output throughput than Gemma 4 31B TP1 under identical hardware: 112 vs 81 output tokens per second at low concurrency, 39 vs 30 TOPS at high concurrency (Cohere blog, 2026-09-10).

Sovereign-AI framing. Cohere positions this as part of its sovereign-AI push — a self-hostable translation model for governments and regulated buyers. The CC BY-NC 4.0 license restricts commercial use; commercial deployment goes through RWS Language Weaver or Cohere’s sales team.

Architecture

Decoder-only sparse MoE Transformer: 218B total / 25B active, 128 experts (8 activated per token) with shared experts. Sliding-window attention (window 4096) with RoPE interleaved with global attention in a 3:1 ratio — the same pattern from Command A. 16K input, 16K output, text → text (HF model card).

The router uses sigmoid activation over expert logits, normalized over top-k. All three quantization checkpoints (BF16, FP8, NVFP4 W4A16) are the production weights. Minimum hardware: 1× B200 or 2× H100 at W4A16; up to 4× B200 / 8× H100 at BF16 (HF model card).

Benchmark numbers — and how they were measured

The headline score is 83.60 across all languages on what Cohere labels “WMT26 benchmarks” (Cohere blog, 2026-09-10). But the evaluation method is critical: Cohere ran its own benchmark using GPT-5.6-Sol as a judge (Cohere blog, footnote 1). This is not the official WMT26 shared task.

The official WMT26 General MT task uses human evaluation (“All systems will be human evaluated”) and does not release automatic preliminary results. WMT26 results have not been published as of 2026-09-12 — the conference is 28-29 October 2026 in Budapest. The 83.60 number is Cohere’s own GPT-5.6-Sol-judged run, not an official WMT26 shared-task score.

Cohere also reports a long-context score of 48.9 (measured via xComet-XL on two-chapter book translation), compared to 21.3 for Google Translate and 19.4 for Gemma 4 31B (Cohere blog, 2026-09-10).

Regional averages from the same Cohere run show North Small Translate beating Gemma 4 31B in Europe (82.2 vs 73.9) and running roughly even in South Asia (86.2 vs 86.7). Against DeepL NextGen, the largest gaps are in South Asia and MENA (~8-10 points), with the narrowest edge in East Asia (~1-3 points).

Practical implications

• Self-hostable. Weights are on Hugging Face under CC BY-NC 4.0. Three quantization tiers are available (BF16, FP8, NVFP4 W4A16). The minimum is 1× B200 or 2× H100 at W4A16.
• Production inference. Cohere recommends greedy decoding (matching its production setup). The model can run via transformers, vLLM, or the Hugging Face Space demo. Cohere’s melody library is needed for response parsing.
• Commercial path. CC BY-NC 4.0 means research and non-commercial use only. Enterprise deployments go through RWS Language Weaver or Cohere sales.
• 50 languages. The model covers 32 high-resource and 18 additional languages, with no sharp regional drop-offs that smaller models typically show.

Risks and caveats

• Vendor-run benchmark. The 83.60 score, throughput, cost, and regional breakdowns are all Cohere-reported. No independent verification exists.
• GPT-5.6-Sol as judge. The headline number depends on the judge model’s calibration, not human evaluation. The official WMT26 shared task uses human evaluation and has not published results — the conference is 28-29 October 2026.
• CC BY-NC 4.0. Non-commercial use only. Commercial deployment requires RWS Language Weaver or Cohere sales.
• Tom Kocmi dual role. A Cohere employee organizes WMT26 (WMT26 page) — relevant context, not an endorsement of the self-run benchmark.

What to watch

• Official WMT26 human-evaluated results (October 2026)
• Independent reproductions of the Cohere-run benchmark
• CC BY-NC license impact on adoption vs permissively licensed alternatives
• RWS Language Weaver enterprise uptake

Sources