Command A is an enterprise-focused large language model (LLM) developed by Cohere, designed for high-performance agentic tasks and retrieval-augmented generation (RAG). Released in March 2025, it serves as a successor to the Command R series, emphasizing inference efficiency and hardware accessibility for private enterprise deployments.
The model features 111 billion parameters and supports an extensive 256,000-token context window. A primary design goal was to minimize hardware requirements; Command A is optimized to run on as few as two GPUs (such as NVIDIA A100 or H100), offering significantly higher throughput compared to its predecessor, Command R+. Its architecture is built for "max performance, minimal compute," catering to organizations with strict data security and cost-efficiency requirements.
Command A is natively capable of complex tool use and iterative reasoning, allowing it to function as a core engine for AI agents that interact with external APIs, databases, and search engines. It supports verifiable citations in RAG workflows to improve the reliability of generated information by sourcing answers directly from provided documents.
The model is trained across 23 business-critical languages, including English, French, Spanish, Japanese, Chinese, and Arabic. While primarily an enterprise product, Cohere released the weights for research purposes under a CC-BY-NC license through its Cohere Labs initiative.