Mistral's Ministral 3B is a small, multimodal AI model built for edge and resource-constrained deployment, able to run on a single GPU in roughly 8 GB of memory, or less when compressed. "On-device" AI runs on local hardware such as a laptop or a shop computer rather than sending data to a cloud service.
Running a model locally keeps data on the premises and works without an internet connection, at the cost of some peak capability compared with the largest cloud models.
What it means: for tasks that touch sensitive customer or health data, a small model running on your own hardware can be both cheaper and more private than a cloud service. It will not match the biggest models, but for classifying, summarising and extracting text it is often more than enough.