Cohere Labs published the weights for North Small Translate, a sparse mixture-of-experts model built specifically for machine translation. The model activates 25 billion of its 218 billion total parameters per token, routing each request through 8 of 128 experts plus a set of shared experts that process every token regardless of routing. That is a research release, not a product launch, and the distinction matters for anyone hoping to run it in production.
The architecture borrows a detail from Cohere’s Command A: attention layers alternate between sliding-window blocks (a 4096-token window using rotary position embeddings) and global attention layers with no positional embeddings at all, in a 3 to 1 ratio. Rather than the more common softmax, Cohere Labs says its router passes expert logits through a sigmoid, normalizing only across the selected top-k experts. The model handles 16,000 tokens of input and 16,000 of output, and covers 50 languages spanning European, Middle Eastern, South Asian, and East Asian scripts, from Icelandic and Maltese to Tamil, Telugu, and both Traditional and Simplified Chinese.
On Cohere’s own WMT26 evaluation, the model scores 83.60 across all tested languages. Running an agentic multi-pass translation workflow instead of a single forward pass lifts that to 84.36. Cohere Labs has not published how a comparable general-purpose frontier model scores on the same suite in this release, so the number stands as a claim about the model’s own ceiling rather than a head-to-head result against a named competitor.
The licensing term is the part worth reading past the model card’s specs table. North Small Translate ships under a CC BY-NC 4.0 license, meaning noncommercial use only, layered with Cohere Labs’ own Acceptable Use Policy. The terms-of-use section states this directly: researchers get open weights, but any commercial deployment requires contacting Cohere’s sales team for a separate license. That is a meaningfully different offer than a permissive open-weight release like Mistral’s Apache-licensed models, where a startup can ship a translation feature in a paid product the same day it downloads the weights.
For a translation model, that restriction lands squarely on the buyers most likely to want it. Enterprises with high-volume localization needs, the exact audience Cohere pitches its production API to, cannot legally deploy North Small Translate against paying customers without a separate commercial agreement. The open-weights release functions as a research artifact and a marketing signal of Cohere’s translation capability, while the actual revenue-generating deployment path runs through Cohere’s hosted product.
Running the model also demands real infrastructure. Cohere Labs lists 8 x H100 GPUs as the minimum for a full 16-bit checkpoint, dropping to 4 x H100 at 8-bit precision and 2 x H100 for a 4-bit quantized variant. All three quantization levels are described as the same checkpoints Cohere runs in its own production stack, which suggests the quantized versions were validated rather than bolted on as an afterthought. The model card also flags a specific deployment trap: loading with device_map="auto" alone can exhaust memory mid-load, because it fills every GPU with weights before leaving room for the temporary buffers that fuse the mixture-of-experts layers, so Cohere Labs recommends capping per-device memory explicitly during loading.
The release does not include a licensing cost for commercial use, so enterprises weighing it against production-ready alternatives cannot yet run the pricing comparison that would determine whether it beats a paid API on cost per translated word. Teams evaluating open-weight translation stacks for regulated or high-volume workloads should treat North Small Translate as a benchmark to test against internally, not a deployment candidate, until Cohere publishes commercial terms.
According to the model card published by Cohere Labs on Hugging Face on September 10, 2026.
