Compatibility and requirements

IndeRun has two kinds of requirements. The install floor is what lets your app build and start. The model requirement is what lets a device run a task on its own model. A device can meet the first and not the second, which is why IndeRun checks at runtime and refuses when no eligible model is there.

Tested set: IndeRun 0.5.0-rc.17 (@independo/inderun-web, SwiftPM, Maven) and @independo/capacitor-inderun 1.2.0, which wraps IndeRun 0.4.0. Examples type-checked against these on 2026-11-10.

Supported platforms and limitations

You build with	Package (tested set)	Installs on	On-device model, checked at runtime
Capacitor 8	@independo/capacitor-inderun 1.2.0 (wraps IndeRun 0.4.0)	iOS 16+, Android 8.0 (API 26)+, modern browsers	iOS 26+ on an Apple Intelligence device with it turned on; Android devices with AICore/Gemini Nano; Chrome Prompt API on desktop Chrome when systemModel is configured
Web / TypeScript	@independo/inderun-web 0.5.0-rc.17	ES2022 + WebAssembly (Node 24+ for tooling)	Chrome Prompt API; ONNX Runtime Web with a model you supply
Swift	IndeRun 0.5.0-rc.17 (SwiftPM)	iOS 16+, macOS 14+, Swift 5.9+	Apple Foundation Models (iOS/macOS 26+)
Kotlin	app.independo.inderun:inderun-kotlin 0.5.0-rc.17	Android 8.0 (API 26)+, JDK 17	ML Kit GenAI / Gemini Nano

Versions to install:

You build with	Install (tested set)
Capacitor 8	npm install @independo/capacitor-inderun@1.2.0
Web / TypeScript	pnpm add @independo/inderun-web@0.5.0-rc.17
Swift (SwiftPM)	.package(url: "https://github.com/independo-gmbh/inderun.git", exact: "0.5.0-rc.17")
Kotlin (Maven Central)	implementation("app.independo.inderun:inderun-kotlin:0.5.0-rc.17")

Capacitor toolchain: Capacitor 8, Xcode 26, compileSdk 37, Android Gradle plugin 9.1+, Kotlin Gradle plugin 2.4.10, iOS deployment target 16.0. The exact settings and failure strings are in the Capacitor quickstart and the plugin’s Host Project Requirements (opens in new tab).

Kotlin SDK host requirements (compileSdk, Android Gradle plugin, Gradle, Kotlin Gradle plugin, JDK), each with the build error it prevents and the releases it applies to: Android host requirements (opens in new tab).

The OpenAI-compatible cloud provider is available on every platform, but it is a route you configure and allow; it is never used for local_required requests.

Usage guides for each SDK: Web (opens in new tab), Swift (opens in new tab), Kotlin (opens in new tab).

Honest limitations

• Whether an on-device model is available depends on the device, its settings and the model’s state. IndeRun reports it; it cannot create it.
• No claim is made that tasks work offline in general. A model that is not downloaded or not ready is unavailable.
• There is no edge provider today, and IndeRun does not fall back to a cloud provider automatically: a cloud route exists only if your app configured one and the request’s rule permits it.
• Realtime sessions (openSession()) are not implemented. run() and stream() are; not every provider streams.
• The Capacitor plugin release listed above has no speech API. Speech is available through the Web, Swift and Kotlin SDKs.
• IndeRun is an execution layer. It is not a model, and not an accessibility certification.

What each platform needs, checked in one call

Platform	On-device model	Why it may be unavailable
iOS	Apple Foundation Models	Device not eligible, Apple Intelligence turned off, model not ready yet. See Apple’s availability reasons (opens in new tab).
Android	Gemini Nano via ML Kit GenAI	Device unsupported by AICore (the API is also not supported with an unlocked bootloader), Gemini Nano not downloaded, app not in the foreground, per-app inference quota. See ML Kit GenAI overview (opens in new tab) and Prompt API: get started (opens in new tab).
Web	Browser-managed model (Chrome Prompt API)	Chrome documents the Prompt API for desktop Chrome, with OS, storage and hardware requirements; other browsers don’t implement it yet. Model download pending or unavailable. See Chrome’s Prompt API (opens in new tab).

IndeRun turns these into one capability check and one routing decision. See When on-device AI is unavailable.

Speech

Speech is a separate capability with its own rules; see Read text aloud.

Platform	Playback (speak())	Returned audio (generateSpeech())	How locality is decided
iOS	Yes, apple_system_speech	WAV	Installed Apple voices are assumed local, because the API exposes no network flag (the airplane-mode check is pending)
Android	Yes, android_system_speech	WAV	Enforced per voice through its network flag (Voice.isNetworkConnectionRequired); an engine that lists no voices is reported as voice_locality_unknown
Web	Yes, web_system_speech	Refused: playback only	Browsers can’t confirm locality, so local_required speech is always refused on the web

A missing voice fails rather than being substituted. No offline-speech claim is made.

Device records

Observed results on real devices are kept in the repository, not on this page: see docs/verification/ (opens in new tab). That folder is on GitHub only.