What is the main reason you use local AI models today: cost, privacy, speed, or avoiding usage…

My short answer: people choose local models for more than price.

The strongest reasons are privacy and data sovereignty, offline operation, predictable latency, and control over the model and software stack. A local model can be useful when sensitive text must not leave the device, when connectivity is unreliable, or when a stable workload makes hardware ownership worthwhile.

But “local” does not automatically mean cheaper or greener. A 2026 study of 18 model configurations on two smartphones and a server found that on-device inference was, on average, about three times less energy-efficient than batched server inference; the authors also estimated that 88–90% of the modeled environmental impact came from device manufacturing rather than electricity. See the study.

So my rule would be:

• choose local first for private, offline, low-latency, or highly repetitive tasks;
• choose an API when you need frontier capability, burst capacity, or a long context;
• compare total cost, including hardware, electricity, maintenance, and your time.

The interesting question is not “local or cloud?” but “which parts of this workflow must stay local, and which parts benefit enough from a stronger remote model to justify the trade-off?”

From what we see with Linux desktop users: privacy comes first (mail, notes and files stay on the machine), then predictable cost and no usage caps. Speed is usually the trade-off rather than the reason, though a 7-8B model at Q4 on a mid-range GPU is plenty for drafting, summarising and quick questions. The underrated one is control: you pick the model, and it doesn't change under you.