Cohere Launches Open-Weight North Small Translate
Cohere launched North Small Translate, an open-weight 218B translation model. Here are the benchmark scores, hardware limits, and commercial terms.
The 218-billion-parameter system targets enterprise data privacy, but requires serious GPU hardware and commercial licensing.
Cohere released North Small Translate on September 10, 2026, marking the Canadian firm’s first dedicated model for machine translation. The company published model weights on Hugging Face under the non-commercial CC BY-NC 4.0 license. The release targets enterprises that need to process internal legal contracts and sensitive documents without routing text through third-party cloud translation APIs.
First: How North Small Translate Works
The model uses open weights, meaning researchers can download and inspect the files locally for testing. It is not open source in the traditional sense, as the license explicitly bars commercial production use without a paid agreement.
Under the hood, the system relies on a sparse Mixture-of-Experts (MoE) architecture. Instead of running every parameter for each translated word, the network divides its workload across 128 smaller expert blocks. For each generated token, the router activates only 8 expert blocks alongside shared modules. While the model contains 218 billion parameters in total, only 25 billion are active during any individual token pass. That routing keeps inference computing requirements manageable while preserving a deep vocabulary. The training was completed in technical partnership with RWS’s Language Weaver division for business terminology tuning. Cohere did not disclose the total volume or exact composition of the underlying dataset.
Benchmark Numbers: Promise Versus Reality
In academic evaluations on the WMT26 benchmark, Cohere reported an average score of 83.60 points in direct translation and 84.36 points when using an agentic multi-pass workflow with automated error correction. According to the company, the system outperformed established translation engines across more than 50 languages.
Those numbers call for critical context. All published scores were measured internally by Cohere rather than verified by an independent evaluation team in production environments. While promotional materials focused on the wide gap against Google Translate (which scored 68.20), the comparison against DeepL NextGen tells a different story: DeepL reached 81.37, leaving Cohere with a modest 2.23-point lead.
For lengthy documents, the engine supports 16,000 input tokens and 16,000 output tokens. In Cohere’s internal tests translating two-chapter book excerpts measured by xComet-XL, the model logged 48.9 points, compared to 21.3 for Google Translate and 19.4 for Gemma 4 31B.
Hardware Footprint and Commercial Terms
Running the model on private infrastructure requires substantial enterprise hardware. The smallest quantized variant (NVFP4 W4A16) needs at least one Nvidia B200 GPU or two Nvidia H100 cards. Deploying uncompressed precision variants like FP8 or BF16 demands between two and eight high-end accelerators.
The free download on Hugging Face covers only research and academic testing. Organizations looking to run North Small Translate in commercial environments must negotiate deployment through Cohere Model Vault or query Cohere’s commercial API. The company estimates standard API inference costs at roughly $0.000676 per task of 661 tokens.
Who It Serves and What Remains Open
The model suits research teams and large enterprises that already operate dedicated GPU clusters and must maintain strict on-premises data isolation. It does not make sense for small teams lacking dedicated AI hardware. What remains unproven is real-world performance under independent audits, as well as operational latency during sustained production workloads.