Hardware adaptation
Memory, compute, storage and acceleration are read from the machine. The plan is made for that machine, not for an average one, and it changes when the hardware does.
Technology
How the runtime adapts AI to your machine.
A model that fits on paper does not always run well in practice. The AI Stays Local runtime plans around the hardware that is actually there, recommends only profiles with recorded evidence, and keeps the whole lifecycle on your machine. This page explains the approach, not the engine’s internal methods.
The approach
Seven principles behind every profile.
Model profiles
A profile is a specific open-weight model at a pinned revision and a stated precision, with the hardware it was evaluated on and what the evaluation found. It is the unit the runtime recommends, installs and runs.
Memory, compute and storage awareness
Usable memory is total memory minus what the operating system keeps. Unified memory, system RAM and discrete GPU memory are treated as different resources, and free storage is part of the plan.
Local operation
Inference runs on the machine you control. Normal local inference is not billed by the token, and a question is answered where it was asked.
Execution lifecycle
Acquisition, verification, startup, health, stopping and removal belong to the runtime. A failed installation does not replace a verified model that was already working.
Evidence-based validation
Running, being validated, being usable and being available are separate facts, recorded separately. A parameter count is never treated as evidence, and every published figure names its hardware and method.
A clear privacy boundary
Prompts and documents are processed on the machine during local inference. The network is used to obtain the runtime and model files, licence checks carry no content, and telemetry is off by default.
Why profiles, not models
The same model can be the right choice on one machine and the wrong one on another.
Whether a model is useful depends on the model, its precision, the memory it can use, the acceleration available, the speed of the storage it is read from and the work it is asked to do. A memory figure on its own predicts very little.
So the runtime recommends profiles rather than model names. Each profile carries the evidence behind it and a state for each dimension, and the recommendation for a machine is drawn only from profiles whose evidence covers that class of hardware.
What is published
What you can check, and what stays inside the product.
- The profile: model, publisher, licence, pinned revision and precision.
- The hardware each measurement was taken on, and the method behind it.
- The state of each profile: whether it runs, is validated, is usable and is available.
- How acquisition is verified, and what the local API accepts.
How the engine places and moves model data in memory is part of the product and is not described on this site.
Credit where it is due
Built on open models, credited by name.
Every model belongs to its publisher, under its own licence. Upstream engine work used by the runtime is credited in the product notices, and results reported by upstream projects are never presented as ours. AI Stays Local is the product layer on top: hardware planning, acquisition, runtime orchestration, lifecycle management and the applications. The runtime is proprietary and is not described as open source.