Cohere North Small Translate: MoE Translation Model Debuts

What Cohere Released

Cohere has released North Small Translate (model name North-Small-Translate-1.0), a mixture-of-experts machine translation model. According to Cohere, it is the first translation model in the North model family, built on the company's lineage from the Tiny Aya model family and Command A Translate. The release targets translation quality and cost efficiency at scale, rather than functioning as a general-purpose chat model.

Specialized translation models matter because general-purpose large language models often carry overhead unrelated to translation, while dedicated systems can struggle with context length or language coverage. Cohere positions North Small Translate as an alternative that combines broad language coverage with a mixture-of-experts design meant to keep inference costs down.

Architecture and Feature Overview

North Small Translate uses a mixture-of-experts (MoE) architecture with 218B total parameters and 25B active parameters per forward pass, according to Cohere. It supports a 16k-token input and 16k-token output context window, works purely in text in / text out mode, and covers more than 50 languages. Cohere describes it as optimized specifically for machine translation rather than general instruction-following.

Running the model requires meaningful hardware: Cohere lists a minimum configuration of one Nvidia B200 GPU or two H100 GPUs, both at W4A4 quantization. This reflects the model's overall size even though only a fraction of parameters activate per token.

Alongside the standard model, Cohere also evaluated an "Agentic" variant designed to find and fix translation errors before returning output. This variant is scored separately in Cohere's benchmark results and consistently outperforms the standard model.

North Small Translate is released as open weights under a CC BY-NC 4.0 license, restricting use to research and non-commercial purposes. Cohere has published the weights on Hugging Face under CohereLabs/North-Small-Translate-1.0, alongside several near-lossless quantized versions, a demo Space, and implementation documentation. Enterprises seeking a commercial license can access the model through RWS's Language Weaver product, described below.

Usability and Performance Analysis

According to Cohere's benchmarks on WMT26 (all-languages, judged with GPT-5.6-Sol as the LLM judge), North Small Translate and its Agentic variant scored higher than several comparison models:

Under WMT scoring bands, scores in the 80-100 range indicate perfect or minor-error translations, while 60-80 indicates good translations with major errors. By this scale, Cohere's reported scores for both North Small Translate variants sit in the higher band, while Google Translate's reported score falls into the "good with major errors" range.

Cohere also reports regional breakdowns across 32 high-resource languages plus 18 additional languages, describing North Small Translate as the most consistent performer in its comparison set, without the sharp regional drop-offs it observed in similarly sized models. In Europe, Cohere reports 82.2 for North Small Translate against 73.9 for Gemma 4 31B (on); in South Asia, the two are close, at 86.2 versus 86.7. Broken out by variant, Cohere's Agentic and standard scores beat Gemma 4 31B (on) in EU languages (82.74 / 82.17 vs. 72.73) and non-EU European languages (81.52 / 81.24 vs. 75.90), while Gemma leads narrowly in South Asia (88.04 vs. 87.13 / 86.16). Cohere also states both North Small Translate variants outperform DeepL NextGen in every non-European region it tested — MENA, South Asia, Southeast Asia, and East Asia — with the largest gap in South Asia and MENA (roughly 8-10 points), a moderate gap in Southeast Asia (roughly 4-5 points), and the narrowest gap in East Asia (roughly 1-3 points, where DeepL NextGen scored 85.41).

In Cohere's internal throughput tests, North Small Translate produced up to 1.4x higher output throughput than Gemma 4 31B TP1 at identical concurrency and hardware — 112 versus 81 output tokens per second at low concurrency, and 39 versus 30 at high concurrency, roughly a 30-38% gain in tokens per second.

For long-context translation, Cohere's own evaluation translated two chapters of a book in a single call and scored the output per paragraph using xComet-XL. North Small Translate scored 48.9, compared to 21.3 for Google Translate and 19.4 for Gemma 4 31B.

On cost, Cohere reports North Small Translate reaching an 80.1 score at $0.000676 per task, using roughly 661 tokens on average. For comparison, Cohere cites Gemini 3.1 Pro Preview (high) at $0.038928 per task, Qwen 3.5 397B A17B at $0.004525, and its own Command A+ at $0.005158.

Pros and Cons

Pros, based on Cohere's reported figures, include competitive WMT26 scores across a wide language set, regional consistency without sharp performance drop-offs, an Agentic variant that further improves accuracy, meaningfully lower cost per task than several comparison models, and stronger long-context translation quality in Cohere's own book-chapter test.

The clearest limitations are licensing and verification. The CC BY-NC 4.0 license restricts the open weights to research and non-commercial use, so production deployments require a separate commercial license through RWS. All benchmark, throughput, cost, and long-context figures come from Cohere's own testing, with WMT26 scores judged by an LLM (GPT-5.6-Sol) rather than independently verified. The "Small" naming also understates the model's footprint: 218B total parameters require a minimum of one B200 or two H100 GPUs even at W4A4 quantization. Its 16k input / 16k output context window is also modest for document-scale translation work, despite the long-context evaluation targeting book-length chapters.

Outlook

North Small Translate was developed with RWS, involving the company's Language Weaver research and science teams along with language experts. This partnership gives Cohere a path to commercial distribution: enterprises that need more than open-weight research access can obtain the model through RWS's Language Weaver product rather than under the CC BY-NC 4.0 license.

This open-weight-but-non-commercial structure is likely to shape adoption along two tracks. Researchers, academics, and non-commercial developers can download the weights directly from Hugging Face, experiment with the quantized versions, and test the demo Space without cost. Commercial users, however, must negotiate licensing through RWS, which positions Language Weaver as the primary route for enterprise translation deployments built on this model.

Conclusion

North Small Translate gives Cohere a dedicated entry in machine translation, backed by benchmark, throughput, and cost figures the company has published itself. Its regional consistency and cost-per-task figures, as reported by Cohere, stand out relative to the comparison models it tested. The model is best suited to researchers and non-commercial developers evaluating translation quality today, and to enterprises willing to pursue a commercial license through RWS for production use.

Editor's Verdict

Cohere North Small Translate: MoE Translation Model Debuts earns a solid recommendation within the Other LLM space.

The strongest case for paying attention: strong WMT26 scores across a broad language set, according to Cohere's own benchmarks. That alone raises the bar for what readers should expect in this space. Reinforcing that, regional consistency, with no sharp performance drop-offs across Europe, South Asia, MENA, and East Asia — practical value rather than just headline appeal. The broader signal worth registering is straightforward: Cohere's WMT26 benchmarks show North Small Translate's Agentic variant scoring 84.36, ahead of Qwen 3.5 397B A17B (81.56) and DeepL NextGen (81.37). On the other side of the ledger, one constraint is real rather than a marketing footnote: CC BY-NC 4.0 license restricts the open weights to research and non-commercial use; production use requires a commercial license via RWS. It should factor into any serious decision. Layered on top of that, all benchmark, throughput, cost, and long-context figures come from Cohere's own testing, with WMT26 scores judged by an LLM rather than independently verified — which narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.