Local AI is a scheduling problem before it is a model problem

The local-AI conversation tends to focus on the largest model a machine can load. That is interesting, but it is not the question most people will experience.

The practical question is how much useful work the system can do, how quickly, and for how many things at once. Memory headroom, context, caching, concurrency, foreground versus background work, and graceful fallback may matter more than the headline model size.

I would rather have a smaller system that handles ordinary requests immediately and knows when to reach outward than a larger model that technically fits but makes the whole computer feel occupied.

Local AI becomes a product when resource allocation becomes invisible to the person using it. That is a scheduling problem before it is a model problem.