Cohere Command A

Command A, released in March 2025, was for over a year the model Cohere pointed enterprise customers at, and it is still the company's recommended default for anyone migrating off the older Command R line. Its selling point was never raw benchmark leadership: it is a 111-billion-parameter dense model that Cohere states runs on two GPUs, A100s or H100s, with 150 percent higher throughput than the Command R+ refresh it replaced. That combination is the whole product argument. Cohere sells to banks, hospitals and public administrations that deploy inside their own perimeter, and for those buyers the question is how many accelerators a model occupies in a rack, not how it scores on a public leaderboard. Two cards is a number a mid-sized IT department can plan around. The context window is 256,000 tokens with up to 8,000 tokens of output, and the knowledge cutoff is 1 June 2024. The model supports tool use, retrieval-augmented generation with citations, structured outputs, safety modes and multilingual work, and Cohere prices it at 2.50 dollars per million input tokens and 10 dollars per million output tokens. The weights are published on Hugging Face, but under a CC BY-NC 4.0 licence — downloadable and inspectable, yet explicitly barred from commercial use without a separate agreement. That is a different arrangement from the Apache 2.0 licence Cohere later applied to Command A+, and it is worth reading before treating this model as open.

▸Videos

No videos yet.

▸Reviews