Cohere has open-sourced its Command A+ model, a 218 billion parameter sparse mixture-of-experts model, under the Apache 2.0 license. This move aims to facilitate enterprise-grade agent development and private deployment, addressing data sovereignty needs for critical infrastructure. The model activates only 25 billion parameters per inference, optimizing computational efficiency while maintaining performance. It can operate on two NVIDIA H100 GPUs or a single B200 GPU, with Hugging Face offering low-precision quantized versions to ease deployment. Command A+ supports multimodal text and image inputs, featuring a 128K-token input context window and a 64K-token output length. It is optimized for complex logical reasoning, autonomous tool invocation, and long-document processing, supporting 48 languages, including all EU official languages.