Cohere launches Command A, an enterprise model that runs on two GPUs

Background

Large language models are often expensive to run, which limits how many businesses can use them in their own systems. A model that needs fewer processors is cheaper to host, and that matters for companies that want to keep models on their own infrastructure. Cohere, a Canadian enterprise AI company, released Command A in this context.

What happened

On 13 March 2025 Cohere introduced Command A, billed as "max performance, minimal compute"^[1]. The company said the model performed on par with or better than GPT-4o and DeepSeek-V3 on agentic enterprise tasks^[1]. Agentic tasks are jobs in which the model plans and carries out several steps, often with tools.

Key details

The 111-billion-parameter model has a context window of 256k tokens and runs on two A100 or H100 GPUs^[2]. A GPU is a processor widely used to run large AI models. Cohere reported 150% higher throughput than Command R+ 08-2024^[2].

Cohere released the model on its own platform, through its SDK and as non-commercial open weights on Hugging Face^[3][2]. Its SDK is the software kit that developers use to call the model from their own programs.

The model supports 23 languages^[3]. VentureBeat described the model as aimed at global enterprises that need highly multilingual models on minimal hardware^[4].

Impact

The launch framed Command A as a model for organisations that want high performance without large hardware costs. The model's release on an open-weight basis, under a non-commercial licence, also meant that researchers could download and run it while commercial users needed a separate arrangement with Cohere^[2][3].

Limits

Cohere's performance claims are its own, and the comparisons with GPT-4o and DeepSeek-V3 are not independent tests. Throughput and hardware figures describe Cohere's reported setup. Readers should check later evaluations before relying on them for a specific workload.