Command A Reasoning

Cohere’s Command A Reasoning targets complex, enterprise-grade problem solving—especially agentic tool use, retrieval-augmented generation, and multilingual workflows. The model is text-in/text-out with an optional “reasoning” (deliberation) mode that you can toggle on for deeper step-by-step thinking or off for lower latency, without changing models. Specs: 111B parameters, 256K-token context window, and up to 32K output tokens; Cohere notes it can operate on one or two A100/H100 GPUs. An open-weights research release is available via Cohere Labs (HF), including the command-a-reasoning-08-2025 endpoint and templates that expose the reasoning parameter and tool-calling/citations scaffolding. This makes it practical for long-context analysis, audited reasoning traces, and high-throughput agent workflows in production-style environments.