模型發布
Cohere Launches North Small Translate: Open FP8 Weights for a 218B MoE, but Commercial Use Requires a Separate License
North Small Translate activates 25B parameters per token out of 218B total and supports more than 50 languages and regional variants, including Traditional Chinese. The model can be tested through the Chat V2 API or self-hosted using its FP8 weights, but CC BY-NC 4.0 prohibits direct commercial use, and Cohere has not published comparable translation benchmark results.
Cohere released North Small Translate 1.0, a dedicated machine translation model, on September 9, expanding the North family from coding agents into multilingual document processing. The model uses a mixture-of-experts architecture with 218B parameters in total, but activates only 25B per token. Both its context window and maximum output length are 16K tokens. Cohere lists two H100 GPUs or one B200 as the deployment requirement, while the open weights are available only in FP8 format. “Small” therefore refers primarily to the active compute requirement, not to a model footprint that can be easily loaded on a typical workstation.
The language list covers Traditional and Simplified Chinese, Japanese, Korean, English, multiple European languages, and three regional variants of Arabic. Chinese, however, is not among the highest-priority languages. Developers can invoke the model through the existing Chat V2 endpoint using `north-small-translate-1-0`. The interface still requires the source content and target language to be specified in a message rather than introducing a separate translation API. This makes the model easy to integrate into existing generative AI pipelines, but it also means prompt templates, terminology constraints, long-document chunking, and format preservation must still be handled at the application layer.
The release is offered through three channels: a rate-limited free API, FP8 weights on Hugging Face for noncommercial use, and commercially licensed private deployment through Model Vault. This arrangement may appeal to enterprises with stringent data-sovereignty requirements, but “open weights” should not be interpreted as permission to freely incorporate the model into commercial products. CC BY-NC 4.0 and Cohere’s commercial terms must be evaluated separately.
The most significant information gap is currently evaluation. Cohere did not provide WMT, FLORES, COMET, or human-preference results on the release page, nor did it break out performance for Traditional Chinese, specialized terminology, low-resource languages, or long contexts. Engineering teams should first build their own bidirectional language test sets to check numbers, markup, proper nouns, and hallucinatory additions, then compare the model with dedicated translation systems and general-purpose LLMs. On the deployment side, they should benchmark the throughput of its 25B active parameters on the target GPUs and across intended batch sizes and sequence lengths.