Cohere Command A+

Cohere released Command A+ on 20 May 2026 under the Apache 2.0 licence — an unusually permissive choice for a model of this size, and the point of the whole exercise. Cohere sells to banks, governments and health systems, customers whose constraint is not price per token but where the inference happens. A model you can download, inspect and run inside your own perimeter, with no clause restricting commercial use, is a different product from an API key. The architecture is a sparse mixture of experts: 218 billion total parameters with 25 billion active per token, which is what makes the deployment story credible. Cohere states minimum requirements of a single NVIDIA B200 or two H100s with W4A4 quantisation — hardware a mid-sized enterprise can actually buy, rather than a rack it has to rent. Weights are published on Hugging Face in BF16, FP8 and W4A4, and Cohere also offers managed inference through its Model Vault platform. The window is 128,000 tokens in and up to 64,000 tokens out, with text, images and tool calls on the input side and text, reasoning traces and tool calls on the output. It covers 48 languages. Cohere's own figures put the emphasis on agentic work rather than chat. Against its predecessor Command A Reasoning, it reports agentic reasoning on the τ²-Bench Telecom benchmark rising from 37 to 85 percent, coding on Terminal-Bench Hard from 3 to 25 percent, a 20 percent gain in agentic question answering and 32 percent in spreadsheet analysis. Those are step changes rather than increments, and the low starting points are as informative as the new numbers: the previous generation was not usable for this class of task.

▸Videos

No videos yet.

▸Reviews