Cohere launches Command A Reasoning, a language model built for complex enterprise reasoning tasks.
Cohere has introduced Command A Reasoning, a new language model designed for demanding reasoning tasks in business settings. The company says the model performs well in agent workflows, end-to-end systems, and handling large documents. In benchmarks like BFCL-v3, Tau-bench, and DeepResearch Bench, Command A Reasoning reportedly outperforms other private models such as gpt-oss-120b, DeepSeek-R1 0528, and Mistral Magistral Medium.
According to Cohere, the model can run on a single H100 or A100 GPU with a context window of 128,000 tokens, or scale up to 256,000 tokens across multiple GPUs. Enterprises can set a token budget to control resource use and costs, making it possible to run different workloads in parallel without switching models.
Command A Reasoning is available as a research release with open weights under a CC-BY-NC-4.0 license for non-commercial use. Commercial deployments require contacting Cohere's sales team.
Strong performance in deep research
The model is especially focused on complex research tasks. It can power a hierarchical multi-agent system that breaks down requests into subtasks, runs parallel research, and compiles results into structured reports. In related benchmarks, this system achieved top scores for instruction following and content depth.
For safety, Cohere uses an internal evaluation framework to balance strict abuse prevention - covering areas like child abuse, self-harm, violence, sexual content, and conspiracy theories - with minimizing false rejections of legitimate requests. According to Cohere, Command A Reasoning achieves the best tradeoff between safety and usefulness compared to competing models.
Command A Reasoning is available now on the Cohere platform and on Hugging Face.