NAOMS · building block
Local AI models
Pick and run AI models on the person’s own devices, with other providers optional.
What a person notices
- AI features keep working without an internet connection or a subscription to a cloud AI service.
- The person can see, and choose, which model is actually doing the work, instead of a black box.
- A phone or an older laptop gets pointed at a model it can actually run well, not whatever is largest.
What you get as a developer
Local models is the provider layer underneath every AI feature in NAOMS: pick a model, run it on-device, or fall back to a cloud provider when you choose to.
- A provider registry covers local inference backends and cloud providers behind one interface, so a package calls a model without hardcoding which provider is serving it.
- Device-aware model suggestions recommend models a person’s actual hardware can run well, rather than defaulting to the biggest available model regardless of the device.
- Scanning for available models, approving one, and setting a default are separate, real operations a person or an app can trigger any time, not a one-time setup wizard.
- A pooled backend layer manages concurrent inference requests across whichever backends are configured, so several features can share one running model instead of each spinning up its own.
scan for models: model.scan
get a suggestion: model.recommend
set the default: model.default { modelId }