Cohere Command A Reasoning

Command A Reasoning is Cohere's August 2025 flagship: a 111B-parameter reasoning model built for enterprise agentic work — tool use, RAG, multi-step workflows — with 23 languages and a 256K context window. Its distinctive feature is token budgeting: enterprises set how much reasoning compute each request gets, toggling thinking on for hard tasks or off for latency. Cohere reported it beating gpt-oss-120b, DeepSeek-R1-0528, and Magistral Medium on BFCL-v3, Tau-bench, and DeepResearch Bench, and it runs on as few as 1–2 H100/A100 GPUs.

The honest picture: the research release's open weights carry a CC-BY-NC-4.0 non-commercial license — commercial deployment means talking to Cohere sales, not Apache-style freedom. It's enterprise-positioned and priced accordingly, with less hobbyist ecosystem energy than open-weight community models. Vendor benchmarks should be treated as directional rather than definitive.