The gap between the largest models and ones that run on ordinary hardware has narrowed enough to change what is worth doing locally.
What runs now
A model in the range of seven to fourteen billion parameters runs on a laptop with sufficient memory, or slowly on a plain processor. Larger ones need a dedicated graphics card but not a data centre.
At that size, quality is adequate for classification, extraction, summarisation, translation of common languages, and routine code assistance. It is not adequate for anything requiring deep reasoning or broad factual knowledge.
Why anyone bothers
Data never leaves the machine, which resolves an entire category of concern for anyone handling confidential material.
No usage limits, no per-token cost, no service outage, and no terms of service changing under you.
And latency, for short tasks, can be better than a network round trip.
What it costs
Memory, mostly. The model has to fit, and the difference between a machine with enough and one without is absolute rather than gradual.
Time is the other cost. Someone has to set it up, update it, and figure out which model suits which task. That someone is you.
How to try it
Install one of the local runner applications, pull a mid-sized model, and give it three real tasks from your actual work. Half an hour answers the question of whether it is good enough for what you need.
The likely equilibrium
Hybrid, for most people. Local models handling volume and anything sensitive, a commercial service called for the cases where quality genuinely justifies the cost. That is where serious users have already landed.