The biggest concerns with private deployment are the model's black box nature and supply disruptions for updates, and Cohere's recent promises have addressed these issues effectively.

Cohere officially open-sourced the 218B parameter sparse mixture of experts model Command A+, designed for enterprise private deployment and data sovereignty, avoiding vendor lock-in. Inference activation at 25B requires two H100s or one B200, with low-precision quantization options available. Native multimodal input, 128K context, 64K output, covering 48 languages, suitable for complex reasoning, tool invocation, and long document processing.