Cohere Command A Reasoning

Command A Reasoning, announced on 21 August 2025, was Cohere's first entry into the reasoning category — models that generate an internal chain of deliberation before producing an answer. It shares the parameter count of Command A almost exactly, 111 billion, and the same 256,000-token context, which makes the comparison between them unusually clean: the difference is training and inference behaviour, not size. The cost of that behaviour shows up in hardware. Cohere states that Command A Reasoning is optimised for four H100 GPUs in production, twice what plain Command A needs, with four A100s acceptable for evaluation and trials. For a company whose pitch is private deployment on modest hardware, doubling the accelerator count is a visible trade, and it explains why Cohere kept both models in the catalogue rather than replacing one with the other. The model covers 23 languages and targets agent-based tasks and nuanced problem solving rather than open-ended chat. Its maximum output is 32,000 tokens, which is what a model that writes out its reasoning needs. When Cohere released Command A+ in May 2026, it published direct comparisons against this model — agentic reasoning on the τ²-Bench Telecom benchmark rising from 37 to 85 percent — so the company itself now treats Command A Reasoning as the prior generation. The weights are on Hugging Face under CC BY-NC 4.0, which permits inspection and research but not commercial deployment.

▸Videos

No videos yet.

▸Reviews