Cohere's translation model carries the same 111 billion parameters as its flagship — and a window sixteen times smaller

Cohere's dedicated machine translation model, Command A Translate, is documented at 111 billion parameters with a context length of 16,000 tokens — 8,000 for input and 8,000 for output. The weights published on Hugging Face contain 111,057,580,032 parameters, which is the exact figure carried by Command A, the company's general-purpose enterprise flagship. The identical count points to a specialised derivative rather than a separately trained model. What differs is what the model is allowed to see at once. Command A works with a 256,000-token window; the translation variant works with 16,000. A model of the same size therefore reads sixteen times less in a single pass, which rules out handing it a full contract or a book and pushes long documents into chunked workflows — a constraint Cohere addresses in a separate cookbook for translating long texts. The rest of the specification follows the family. Cohere lists 23 supported languages, including Polish, Arabic, Chinese, Hindi and Persian, and states the model runs on one or two A100 or H100 GPUs — the same hardware argument the company uses to sell Command A to buyers who count accelerators in a rack. Cohere does not list a per-token price for Command A Translate on its public pricing page, which is consistent with a product sold mainly through private deployment agreements: the pitch is translating sensitive material inside the customer's own perimeter. The downloadable weights carry a CC BY-NC 4.0 licence, so they can be inspected but not deployed commercially without a separate agreement.