Command A

Cohere’s Command A (Mar 2025) is a 111B parameter model that delivers GPT-4-tier performance with dramatically better efficiency—requiring just two A100s or H100s to deploy, compared to 8–32 for most competitors. It supports a 256K context window (2× GPT-4o), hybrid attention (sliding + global), and excels across enterprise-critical benchmarks: outperforming or matching GPT-4o and DeepSeek-V3 in human evals for business, coding, and agentic tasks. Notably, it leads on BFCL-v3 and Taubench for tool-using agents, performs on par with GPT-4o on MMLU, MBPPPlus, and SQL, and dominates RepoQA in long-context code understanding. With token throughput up to 156/sec—1.75× GPT-4o, 2.4× DeepSeek-V3—Command A is fast, accurate, and enterprise-ready. But what makes it useful is how well it translates benchmark wins into real-world capabilities. Command A is designed for RAG with citations, structured tool use, and multilingual fluency across 23 languages, including best-in-class dialect control (e.g., Arabic ADI2 scores). It ships with strict/contextual safety modes, runs well in private deployments, and integrates seamlessly into Cohere’s North agent stack for secure internal automation. Its open weights (non-commercial) and cost-effective footprint make it uniquely viable for teams building internal copilots or vertical agents. Among frontier-class LLMs, Command A may not shout the loudest—but it speaks the clearest, and fastest, where it counts.