Apple On-Device AI Model: What Developers Can Access

Updated 10 November 2026 · Standard formula, shown below

Yes. Apple has opened its on-device foundation model to developers through the Foundation Models framework, introduced at WWDC 2025. Eligible apps can call the model on supported Apple devices for tasks such as summarizing, extracting information, and generating structured responses, without sending those requests to a cloud model.

What Apple made available

The Foundation Models framework gives developers access to Apple’s own on-device language model through system APIs. It is part of Apple’s broader Apple Intelligence platform, but it is distinct from a general-purpose hosted chatbot service: the framework is intended for focused language tasks performed locally on the device.

Apple announced the framework in June 2025 and made it available to developers with the Apple operating-system software development kits for its 2025 platforms. Developers can build it into apps rather than distributing or hosting the underlying model themselves. Apple’s developer documentation describes the model as available without a per-request inference charge, though developers still need to account for their ordinary app development and distribution costs.

Access is not the same as unrestricted access to every Apple Intelligence feature. The framework exposes model capabilities through defined interfaces and system safeguards. It does not provide developers with the model’s weights to download, modify, or run on arbitrary hardware.

How the on-device model works

A developer integrates the framework into an app and uses its model session and generation APIs to submit a prompt or a more structured request. The model then produces a response on the device when the necessary Apple Intelligence support is available. Local processing can reduce the need to transmit a user’s text to a server and may allow features to work without an internet connection, subject to the device, operating system, and task.

The framework is designed for common language capabilities, including:

Generating or rewriting short text, such as a suggested reply or a concise description.
Summarizing text the user has provided or content the app is permitted to process.
Extracting information into a defined format, such as identifying dates or categories in a note.
Classifying text or returning structured output that an app can use to guide a workflow.

Developers can also describe expected output using structured generation features, which can help an app request a particular data shape rather than treating every response as free-form prose. The result still needs appropriate validation: a model can misunderstand a prompt, omit information, or produce an unsuitable answer.

Which devices and users can use it

“On-device” does not mean that the model will run on every iPhone, iPad, or Mac that can install an app. It depends on Apple Intelligence being supported and available on the device, as well as the required operating-system version, language, region, and user settings. Apple’s supported-device requirements for Apple Intelligence include newer hardware with sufficient capability; developers should check the current platform documentation instead of assuming compatibility from an app’s minimum operating-system version.

An app can therefore be available on devices where the Foundation Models framework cannot produce a response. Developers need to check model availability at runtime and provide a sensible alternative, such as a non-AI workflow, a simpler local method, or—if the app offers one and the user agrees—a separate online service. An online fallback changes the privacy and connectivity assumptions, so it should not be treated as equivalent to local processing.

What developers can and cannot control

Developers control the app experience: what task to offer, what context to provide, how to present the result, and whether to ask the user to review or edit it. They can shape prompts and use supported generation controls, but they do not control the model’s internal weights or turn it into an unrestricted model-training platform.

Apple also provides mechanisms intended to help keep outputs within the requested format and to handle refusal or unavailability. These protections do not guarantee factual accuracy or make every proposed use appropriate. Apps should avoid implying that generated text is verified, especially where an incorrect result could have meaningful consequences. A useful design is to show the source material alongside a summary or to make extracted fields easy to correct.

Local inference can offer a privacy advantage because the model request need not be sent to a remote model provider. That does not automatically make an entire app private: the app may separately collect analytics, sync user content, or transmit data for other reasons. Developers should explain their app’s actual data flows and request only the information needed for the task.

When it makes sense—and when it does not

The framework is a good fit when a feature is narrow, text-focused, and useful even if the response is brief or imperfect. For example, a notes app could offer a short summary of a selected note, while a mail app could suggest a draft reply for the user to review. A catalog app might classify a user-written description into a small set of categories.

It is less suitable when an app needs up-to-date facts from the web, very long or complex reasoning, guaranteed exact answers, or large-scale processing beyond what the device can handle. The model’s knowledge is not a live search index, and local generation should not be presented as a substitute for authoritative current data. For those needs, an app may combine the local model with search, a database, or a cloud service—but each source and its limitations should be made clear.

Developers should test practical concerns such as response latency, output quality across supported languages, behavior with long input, and what users see when the model is unavailable. They should also avoid sending an entire conversation or document when a smaller relevant excerpt will do. Keeping prompts focused can improve both usefulness and resource efficiency.

Frequently Asked Questions

Did Apple release its model for developers to download?
No. Apple provides access through the Foundation Models framework. Developers call the model through supported system APIs; Apple has not made the model’s weights available as a downloadable package for arbitrary use.

Does the framework require an internet connection?
The model is designed to run on-device, so a model request can be processed locally when the device supports the feature and the model is available. Other parts of an app may still require connectivity.

Is it available on every Apple device?
No. Availability depends on supported hardware and software, Apple Intelligence availability, and relevant language or region settings. Apps should check support rather than assume all users can run the model.

Can developers use it without paying for each model request?
Apple presented the framework as having no per-request inference fee. That does not remove the usual costs of building, maintaining, and distributing an app, or costs associated with any separate services it uses.

Can the model’s answers be trusted as facts?
Not without verification. Like other generative models, it can make mistakes or leave out important details. Apps should validate structured results where possible and let users review content before relying on it.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.