Concept
Local-first inference
Running the model, the data and the failure modes on hardware you control.
Local-first is not a deployment preference, it is a claim about who can stop your system. If inference runs on someone else's hardware, then availability, cost, acceptable use and observability are all someone else's decisions.
The engineering consequence is that the constraints become yours to discover. Memory is finite, so one model variant fits at a time. Weights are large, so switching variants mid-batch is a process restart, not a cache miss. The GPU is a single device, so generation is serialised whether your code is concurrent or not.
None of that is a limitation of local inference specifically. It is what inference actually costs, made visible because nobody is absorbing it for you.
Put plainly: understanding you cannot run without permission is not yours. Local-first is how the whole loop — question, model, evidence, answer — stays under one owner.
Implementation
Where this runs
A concept that never left a document is a hypothesis. These are the systems that implement it.