Apple Foundation Models allows iOS developers to integrate on-device AI into an app: summarization, information extraction, text generation, dialogue, text/image analysis, and calls to app functions. For a project, the benefit is clear: less server dependency, better privacy, but compatibility limited to Apple Intelligence devices and essential technical framing.
Apple Foundation Models: what it really is
Announced at WWDC25 on June 9, 2025, Apple Foundation Models is a native Swift framework, meaning a toolbox that Apple app developers can use directly. It provides access to Apple Intelligence models running on the device, without systematically sending data to a server.
Apple presents it as an API for understanding language, generating text, producing structured responses, and triggering actions through tools. Behind that technical wording, the key idea is simple: an iOS app can now ask a local AI model to read, classify, reformulate, or prepare an action, then retrieve a result that the code can use.
Apple Intelligence’s built-in language model was described by Apple Machine Learning Research in 2025 as having around 3 billion parameters. It is not a local GPT-4 in your pocket. It is rather a compact model, designed for latency, privacy, and targeted tasks.
Since iOS 26, iPadOS 26, and macOS 26, Apple indicates that developers can deliver features based on this local large language model. In 2026, the framework was expanded with the protocol LanguageModel, which also allows server models like Claude or Gemini to be connected, according to Apple documentation.
What iOS developers can create with Apple Foundation Models
The strongest use cases are not generic chatbots. They are often much more concrete: saving a user two minutes, avoiding manual entry, reducing a business error. That is where on-device AI becomes interesting for an SMB.
Apple notably cites summarization, entity extraction, text improrvement, game dialogue, creative content generation, and text/image understanding. Apps named by Apple in September 2025, such as SmartGym, Stoic, VLLO, and Detail: AI Video Editor, illustrate this logic: assisting a specific task within an existing experience.
- Internal summary : condensing a meeting note, a customer comment, or a short transcript.
- Information extraction : identifying a date, an amount, a company name, or an address in free text.
- Reformulation : transforming a rough draft into a clearer message, without leaving the app.
- Business assistant : suggesting a next action, for example creating a task or classifying a request.
- Multimodal experience : combining text and image, notably with Vision tools such as OCR (text recognition) or barcode readers documented in 2026.
For a mobile application for businesses, this type of function can replace some of the heavy formulaires. If you compare development approaches before launching a product, this panorama of tools for creating a mobile application helps position Swift, cross-platform frameworks, and the initial trade-offs.
Privacy, server costs, timelines: what it changes for your project
The first impact is privacy. A local AI processes data on the compatible iPhone, iPad, or Mac, which reduces transfers to the cloud. For sales notes, HR drafts, or non-critical health data, this is a real advantage, even if the 2018 GDPR still needs to be handled seriously: purpose, consent, retention period, and individuals’ rights.
The second impact concerns operating costs. An AI feature based solely on a server API like Claude, Gemini, or OpenAI often leads to a variable bill: tokens (units of processed text), network calls, monitoring, and sometimes storage. With Apple Foundation Models locally, some of these costs disappear, but the design and testing budget does not disappear.
Honestly, on a small budget, it is better to aim for a short, measurable AI function rather than a general conversational assistant. A reliable summary in a business screen is often worth more than a chatbot that answers everything approximately.
| AI approach for iOS app | Indicative project cost in France | Realistic timeline | Point of vigilance |
|---|---|---|---|
| Simple local function with Apple Foundation Models | around €8,000 to €20,000 depending on the existing app | 3 to 6 weeks | Apple Intelligence-compatible devices only |
| Local business assistant with the app’s tools and data | around €20,000 to €50,000 | 6 to 12 weeks | Business testing, guardrails, and error cases |
| Hybrid AI local + server model | around €35,000 to €90,000 | 2 to 4 months | API costs, GDPR, network availability |
| Complete AI assistant connected to the information system | often €80,000 and more | 4 to 9 months | Security, access rights, audit, maintenance |
These orders of magnitude vary greatly depending on the state of the application, the quality of the data, and the level of validation expected. On the projects we lead, we often see the same gap: the AI demonstration is quick, but the production-ready version takes time because unexpected responses, device limitations, and user feedback have to be managed.
The limits to know before budgeting
The first limitation is compatibility. Foundation Models features require Apple Intelligence to be enabled and therefore a compatible device. If your user base includes many older iPhones, a core feature that depends solely on this technology can create a two-tier experience.
The second limitation is context size, that is, the amount of information the model can take into account at one time. In February 2026, Apple added APIs such as tokenCount(for:), contextSize and estimates in Xcode Playground. This is useful for avoiding asking the model to read too much text all at once.
Third limitation: local AI does not magically have access to your business data. It can call Tool, that is, functions the application exposes to the model, to search a local database, query a service, or trigger an action. This mechanism is powerful, but it must be locked down. An AI that can create an order, change a status, or send a message must be controlled.
The subtle trap for a decision-maker is confusing a prototype with a product. A demo that rephrases three sentences in English can be put together quickly. A multilingual feature, tested on thousands of cases, accessible, explainable, and maintainable, does not have the same cost.
Local, cloud, or hybrid: the right trade-off
Local is relevant when the data is sensitive, the task is short, and the expected response is fairly well defined. Summarizing a note, extracting a date, suggesting a label, or generating a help sentence: good use cases. Asking for a long legal analysis or a detailed business strategy: less suitable.
The cloud remains preferable when a larger model, broader knowledge, advanced multimodal power, or strong consistency across long documents is needed. Apple’s 2026 updates point in this direction: Foundation Models can work with any model compliant with the protocol LanguageModel, including Apple Foundation Models, server models, Claude, Gemini, or other providers.
Between the two, hybrid is often the best compromise. The app uses local AI for frequent and confidential tasks, then switches to a server model for complex cases, with consent and logging. From an architecture standpoint, this aligns with thinking around AI agents and their connections to data, a topic related to Model Context Protocol for connecting agents to business tools.
Apple also documented in 2026 access to Private Cloud Compute for the latest Apple Foundation Model, intended for apps with fewer than 2 million first-time downloads App Store in total. This option may interest growing publishers, but it requires tracking permissions, the entitlement com.apple.developer.private-cloud-compute and Apple’s current terms.
How to define a Foundation Models feature without scope creep
Good framing starts with a simple sentence: what decision or action should the AI help with? If the answer fits into a specific feature, the project has a good chance of staying under control. If the answer looks like “an intelligent assistant throughout the entire application,” the budget risk rises quickly.
Apple’s guided generation, with @Generable, GenerationSchema and generated Swift structures, is useful for transforming an AI response into clean data. For example: retrieving an array with a date, an amount, and a category rather than a free-form paragraph. For an executive, this is a product quality point, not a developer detail.
Testing must be planned from the start. Apple mentions Instruments for debugging and profiling, as well as tools like the Foundation Models Python SDK and the CLI fm presented in the WWDC26 resources. These tools do not replace business testing, but they help measure response times, errors, and context consumption.
Security must not be handled as an afterthought. If the app handles personal data or automated decisions, the European framework matters: GDPR, internal documentation, and sometimes analysis in light of the 2024 European AI Act. For SMEs already using models like ChatGPT or Claude, this guide on practical compliance with the AI Act provides good reference points.
Final consideration: the interface. A poorly placed AI function looks like a gimmick. A discreet function, close to the user action, is helpful. Recent iOS developments, especially around the Apple design system, can also influence visual integration; the topic is aborded in this analysis of Liquid Glass and its adoption in digital interfaces.
Defining this type of project upfront helps avoid most unpleasant surprises: device compatibility, the real cost of testing, the choice between local and cloud, compliance obligations. It is often at this stage that an outside perspective saves time, before writing the first line of Swift.
FAQ about Apple Foundation Models
Do Apple Foundation Models work without the Internet?
Yes, for features relying on the local Apple Intelligence model, processing can take place on the device. However, an app can also choose a server model or an external service via the protocol LanguageModel.
Can all iPhones use Apple Foundation Models?
No. The features require Apple Intelligence to be enabled and therefore a device compatible with Apple Intelligence. This is a point to verify before making it a core feature of your application.
Is Apple Foundation Models replacing ChatGPT, Claude, or Gemini?
Not entirely. It handles local, well-defined tasks very well, while cloud models like Claude or Gemini remain useful for longer, more complex requests or those connected to large volumes of data.
How much does it cost to add a local AI feature to an iOS app?
For the French market, a simple feature is often around €8,000 to €20,000, depending on the state of the app and the expected testing. A more advanced business assistant can exceed €50,000.