Local AI, explained
What is a local AI assistant?
Last updated:
A local AI assistant runs the language model on your own device, so ordinary prompts can be answered without a round trip to a hosted AI server.
The four distinctions worth remembering
- A local interface is not enough; the model itself must run on the device.
- Offline inference and optional online tools can coexist in the same app.
- Privacy depends on diagnostics, accounts, syncing, and search—not just model location.
- The best model is the one that fits the device and the task, not simply the largest download.
How a local AI answer is produced
The process is easier to judge when it is broken into concrete steps.
-
Download the model weights
A language model is stored as one or more files. A local assistant must put those files on the device before it can use them offline.
-
Load the model into memory
The app prepares the model using the device's processor, graphics hardware, and available memory. Larger models need more resources.
-
Process the prompt locally
The app turns your words—or supported images and files—into input the model can interpret without sending them to a remote inference service.
-
Generate the reply on the device
The model produces the answer token by token. A connection is unnecessary unless the task invokes a separate online feature such as web search.
Local AI and cloud AI take different routes
A cloud assistant cannot answer until the provider receives the prompt. A local assistant can send the same prompt to a model already loaded on the device.
| What matters | Loci | Cloud AI |
|---|---|---|
| Where the model runs | On the iPhone, iPad, or Mac. | On computers operated by the provider. |
| Ordinary chat path | Prompt → local model → reply. | Prompt → network → provider → network → reply. |
| Works without signal | Yes, after the model is downloaded. | No. Every answer needs a connection. |
| Live information | Requires a separate online search. | May be integrated, but still depends on online services. |
| Practical limit | Device memory, storage, heat, and model size. | Provider limits, pricing, availability, and data policy. |
Local is an architecture, not a magic label
An app can run its main model locally and still use the internet for other reasons. That is not inherently deceptive. The problem begins when those reasons are hidden behind a blanket claim that “nothing ever leaves.”
| Ask this | A strong answer sounds like | Why it matters |
|---|---|---|
| Where does inference happen? | On this device, using a named local model or runtime. | This determines whether ordinary prompts need a hosted AI server. |
| What works after disconnecting? | Core chat and a specific list of supported inputs. | “Offline” should describe usable behavior, not merely an app that opens. |
| What diagnostics are collected? | A precise list, a clear exclusion of content, and an off switch. | Local inference does not prevent separate analytics collection. |
| What happens during web search? | The destination, the exact data sent, and whether conversation context is included. | An online lookup creates a separate privacy boundary. |
| How are models chosen? | By device memory, storage, capabilities, and the intended task. | An unsuitable model can be slow, unstable, or impossible to load. |
How to audit a local AI assistant in five minutes
- Find the model. The app should identify what is downloaded or connected locally.
- Disconnect the device. Start a new conversation, not just a cached follow-up.
- Test the inputs you care about. Text may be local while files, images, or voice follow a different path.
- Read the privacy policy by feature. Look separately for chat, diagnostics, search, downloads, syncing, and support.
- Check the failure behavior. When current information is unavailable, the app should say so rather than invent a live answer.
A useful rule: the offline test proves that inference can happen locally. It does not replace a data-flow disclosure, and a privacy policy does not replace the offline test. You need both.
Where local AI is strongest
Good reasons to choose local
- Private drafting, notes, files, or personal questions.
- Flights, travel, dead zones, and unreliable networks.
- Predictable access without per-message charges.
- Control over which model is installed and used.
- Everyday writing, summarizing, coding, and planning.
Reasons to choose cloud
- Frontier-scale reasoning is more important than privacy.
- The task needs extensive, current web research.
- The device lacks the memory for a capable local model.
- You need a large shared knowledge base or cloud collaboration.
- You accept the provider's account, retention, and pricing terms.
How Loci fits the definition
Loci runs AI models on iPhone, iPad, and Mac. Where available, it can use Apple's built-in foundation model or downloadable open models, including Gemma, Qwen, Llama, and Phi, depending on the device and selected model.
Core chat is local. Optional web search is the clearly separated online feature: the specific query goes directly from the device to DuckDuckGo, without sending the conversation through a Loci server. Optional diagnostics cover technical reliability and performance, never prompt or reply content, and can be turned off.
Questions about local AI assistants
What is a local AI assistant?
A local AI assistant runs its language model on your own device. Prompts can be processed and replies generated without sending ordinary chat to a remote AI server.
Is every local AI assistant completely offline?
No. Local inference can work offline while separate features—such as model downloads, web search, syncing, or diagnostics—still use the internet. Each data flow should be evaluated independently.
Does local AI automatically mean private AI?
Not automatically. Local inference removes the need to send prompts to a hosted model, but an app can still collect other data. A trustworthy product explains diagnostics, online tools, accounts, and storage separately.
Why are local models smaller than cloud models?
They have to fit into the memory and storage of one device. Cloud systems can spread much larger models across specialized servers, while local models are optimized to run within a personal device's limits.
Can local AI search the web?
Yes, if the app offers a separate search feature and the device is online. The search request is no longer fully offline, so the app should disclose where that query goes and what context is included.
How can I tell whether an AI app is actually local?
Download its model, disconnect the device, and start a new conversation. Then inspect the privacy policy for diagnostics, search, sync, and account behavior. The offline test proves local inference; the policy explains the remaining data flows.
Product details
Local by default
Run the assistant where your work lives.
Loci brings local models to iPhone, iPad, and Mac with no account, no subscription, and a clear boundary around optional online features.
