Cohere Command A Translate

Command A Translate, published in August 2025, is Cohere's dedicated machine translation model, covering 23 languages. Its selling point is not novelty — automated translation is one of the oldest applications of machine learning — but the combination of quality with private deployment: the model can be run inside a customer's own perimeter, which matters for the contracts, medical records and legal filings enterprises are least willing to send to a public API. The technical profile is unusual and worth reading carefully. The published weights contain 111,057,580,032 parameters — the exact same count as Command A, the company's general-purpose flagship, which strongly suggests a specialised derivative rather than a model trained from scratch. Yet the context window is 16,000 tokens, split into 8,000 for input and 8,000 for output, against 256,000 tokens on Command A. A model of identical size therefore reads sixteen times less at once, which rules out feeding it a whole book and pushes long documents into chunked workflows. Hardware requirements match the family: Cohere states one or two A100 or H100 GPUs, the same figure it uses to sell Command A to buyers who count accelerators in a rack rather than points on a leaderboard. Cohere does not publish a per-token price for this model on its pricing page, which is consistent with it being sold mainly through private deployment agreements. The weights are on Hugging Face under a CC BY-NC 4.0 licence — downloadable and inspectable, but barred from commercial use without a separate agreement.

▸News

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

8/22/2026 · Cohere documentation

▸Videos

No videos yet.

▸Reviews