"Local AI" isn't a spec: how to tell where an AI feature actually runs

"Local AI" gets used as if it were a product category. In practice it only answers one question: where does the main computation for a given feature happen, on your device or infrastructure you control, or on a provider's servers?

That framing helps avoid a few common mix-ups:

• Local is not the same as offline. An app can run inference on-device and still sync, log or fetch data over the network.
• One app, several paths. A single product can process one operation locally and send another to the cloud.
• "On-device", "self-hosted" and "local" overlap but aren't synonyms. A phone and a company server can both count as local while having very different resources and admin models.
• Privacy depends on the whole data flow, not just where the model weights live.

Hardware still sets the ceiling: model size, memory and connectivity decide what can realistically run on a given machine.

The BytePreview glossary entry goes through how local AI works, its real limits and how to check where a specific function is processed:
https://www.bytepreview.com/en/articulos/local-ai-what-it-is-how-it-works-and-its-real-limits?utm_source=dev&utm_medium=social&utm_campaign=bytepreview