Local AI and cloud AI

Neither option is automatically right for every person or every message. The practical difference is where inference happens, who receives the information needed to do it, and what tradeoffs matter for the account or task.

This page is the comparison. For model selection, context length, quantization, and Mac capacity, see Why Spillway uses local AI.

At a glance

Question	Local AI	Cloud AI
Where does inference run?	On hardware the user controls	On a provider's infrastructure
Does selected content leave the computer?	Normally no	Yes
Internet required for ordinary inference?	Usually no after setup	Yes
Model capability	Limited by local hardware/model choice	Often access to larger frontier models
Operational cost	Local memory, storage, power, and compute	Provider usage/subscription costs may apply
Model control	User chooses what to install/update	Provider controls service/model availability
Private working context known by default	Very little	Very little
Main privacy boundary	The configured local endpoint/device	Provider terms, settings, infrastructure, and account policy

General knowledge and private context are different

A frontier cloud model may know much more about the public world than a model that runs comfortably on a laptop. That can matter.

But neither kind of model automatically knows the user's private working world: current projects, category meanings, rosters, prior corrections, relationships, or months of private email it has never seen.

Spillway therefore treats permitted contextual knowledge as a separate dimension from raw model capability. Deterministic facts, Reference Sources, user-confirmed relationships, and prior decisions can sometimes make a smaller model much more useful on a narrow personal task.

That does not make local models universally better. It means “bigger model” is not the only relevant comparison.

When local AI may fit better

Local processing can be attractive when sensitive material should remain on hardware the user controls, when an organization restricts external processing, when private context should remain local, when offline use matters, or when direct control over model choice matters.

It also consumes the user's own compute resources, and “local” does not automatically mean institutionally permitted.

When cloud AI may fit better

Cloud processing can be attractive when the task benefits from a more capable model, local hardware is insufficient, or the user accepts the provider's terms and data controls for that account.

It may also be faster or less disruptive than running a large model locally.

The choice can differ by account

A personal mailbox, university account, regulated account, and another approved institutional account can reasonably have different AI policies.

Combining those accounts in one workspace should not collapse them into one authorization boundary.

Some knowledge may cross when source content may not

A policy boundary can restrict source material without requiring the system to forget every safe abstraction learned elsewhere.

A user-authored category definition or another appropriately abstracted, provenance-bearing lesson may sometimes be useful across accounts when policy permits, even though restricted message content itself may not cross.

The safeguard is straightforward in principle and demanding in implementation:

Transfer permitted knowledge, not prohibited source content.

Derived fields must not become a back door for reconstructing restricted material. This is an architectural direction, not a claim that every form of cross-account transfer is already implemented.

Local is not automatically permitted; cloud is not automatically forbidden

The relevant boundary depends on the account, organization, law, task, provider, and configuration.

A local model can still violate policy. A cloud provider can sometimes be explicitly approved for certain data. Spillway's goal is to make those distinctions visible rather than encode a universal slogan.

What Spillway should make clear

Before an AI processing option is enabled, Spillway should make clear which service or endpoint will receive the request, whether it is local or remote, what kinds of content may be sent, which accounts may use it, and what cost or policy boundary applies.

The aim is informed choice.

Why Spillway uses local AI · Privacy Policy

---

Documentation provenance: Iterative human–AI construction. See Documentation Provenance.