AI on Device Review (2026): A $4.99 Apple Intelligence-Only App

AI on Device is a native Apple app that runs a single bundled language model directly on an iPhone or Mac, with no server-side component and no internet connection required for inference. It is listed on the Apple App Store under App Store ID 6753206463, developed by Eduardo Dominguez Menendez — an independent developer, not a named company or studio.

The app's own description states it runs a model with "3 billion parameters, fully quantized to 2-bit precision," and claims 100% offline operation with no data collected from the app. It also advertises user-friendly prompt templates and integration built specifically for Apple hardware, rather than a general cross-platform wrapper.

This review evaluates the developer's own App Store claims — the 2-bit quantization figure, the no-data-collection statement, the Apple Intelligence requirement — against what the public App Store listing itself discloses, rather than taking the description at face value. PromptQuorum has not independently benchmarked the model's output quality against other on-device models.

Before paying $4.99, the single most important step is confirming your device actually supports Apple Intelligence — the app will not run otherwise.

AI on Device only works on devices compatible with Apple Intelligence, requiring iOS 26.0 or macOS 26.0 or later — this is the single most important fact for a reader deciding whether the app is usable at all. The App Store listing states this in capital letters as a hard requirement, not a recommendation: the app does not run in a degraded or CPU-only mode on unsupported hardware, it simply does not work.

This requirement excludes a meaningful share of iPhones still in active daily use, since Apple Intelligence compatibility is limited to a specific range of newer iPhone and Mac models. Unlike apps that run a lightweight model on any iOS device with enough RAM (for example, PocketPal AI, which runs any GGUF file a user supplies), AI on Device's single bundled model is tied to the Apple Intelligence platform requirement rather than to a specific RAM or storage floor a reader could check independently.

Practical implication: a reader should verify their specific device against Apple's current Apple Intelligence compatibility list before paying, since the $4.99 purchase price is non-recoverable if the app cannot launch on unsupported hardware.

AI on Device bundles one fixed model — 3 billion parameters, quantized to 2-bit precision — rather than offering a library of swappable models. This is a meaningfully different design from most competitors reviewed in this cluster, which typically let a user download and switch between multiple GGUF, MLX, or curated models.

AI on Device costs $4.99 as a one-time purchase, with no subscription and no usage-based fees, per the current App Store listing.

The App Store listing explicitly frames this as "no fees by use, no quotas of use" — a flat, one-time price rather than a metered or subscription model. This is a straightforward pricing structure compared to apps that combine a free tier with a paid subscription for premium models or features.

The developer states that AI on Device does not collect any data from the app, and that all inference runs 100% offline, on-device.

AI on Device is a straightforward, honestly priced app — $4.99 once, no subscription, no quotas — that runs a fixed 3-billion-parameter, 2-bit-quantized model fully offline. That simplicity is also its biggest limitation: there is no model library to choose from, and the app is gated behind a hard Apple Intelligence hardware requirement (iOS 26.0 / macOS 26.0 or later) that excludes most existing iPhones outright. The honest caveats do not stop there: as of this review the App Store shows too few ratings for a star average to display, and the last update was October 13, 2025 — nearly a year before this review, despite the app being young. That combination — a solo independent developer, no visible track record, and a stale update history — means this is not a purchase to make on reputation. Readers who already own Apple Intelligence-compatible hardware and are comfortable being early adopters of an unproven app can reasonably try it for $4.99. Everyone else, and especially readers who want a broader model library or a longer track record, should start with Private LLM, PocketPal AI, or Noema instead.