Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Apple’s Foundation Models framework gives iOS 26 developers access to an on-device language model for focused app features such as turning a sentence into a workout plan, summarizing a journal, or explaining a user’s own data. Its strengths are privacy, offline-capable inference, and integration with Swift—not broad world knowledge or frontier-level reasoning. It is best treated as a feature component, not a drop-in general-purpose chatbot.
That distinction shapes the design decision: use the system model to interpret or express information your app already has, and use validated app logic or tools for facts, calculations, and actions that must be correct.
What “local AI” means in iOS 26
Apple’s Foundation Models framework is a Swift API for working with the system language model behind Apple Intelligence. The model runs on compatible devices rather than being packaged inside each app or called through a cloud inference API. Apple says the on-device path can work without a network connection and has no per-token cloud API charge; those claims apply to model inference, not necessarily to the rest of an app’s services or operating costs. Apple’s overview of the framework describes its intended use in app experiences.
Local inference can help keep the prompt and response on the device for that model interaction. It does not automatically make an entire feature private: analytics, synchronization, remote search, account services, and tools that contact a server can still transmit data. Nor does “iOS 26” mean every device running that operating system can use Apple Intelligence. The device must be eligible, Apple Intelligence must be enabled, and the model must be ready. Apple’s Foundation Models task documentation describes checking availability before offering the feature.
#1 Best Overall
- This phone is unlocked and compatible with any carrier of choice on GSM and CDMA networks (e.g. AT&T, T-Mobile, Sprint, Verizon, US Cellular, Cricket, Metro, Tracfone, Mint Mobile, etc.).
- Please check with your carrier to verify compatibility.
- The device does not come with headphones or a SIM card. It does include a generic (Mfi certified) charging cable.
- Tested for battery health and guaranteed to have a minimum battery capacity of 80%.
Apple described the iOS 26-era on-device model as approximately three billion parameters, quantized to two bits, and optimized for tasks such as summarization, extraction, and classification. Apple also cautions that it is not designed for broad world knowledge or advanced reasoning. Treat those specifications as applying to the iOS 26-era model—not as a permanent description of every later Apple model generation. Apple’s WWDC25 Foundation Models session explains the target tasks and design.
Five useful product patterns
1. Turn natural language into app data
A user might type, “Make me a 30-minute dumbbell workout,” and receive a draft plan with exercises, sets, repetitions, and rest periods. Similar patterns include turning a note into a title and tags, or converting a search phrase into filters. The important part is not merely generating plausible text: it is producing values the app can inspect, validate, edit, and use.
For this job, Apple’s guided-generation facilities—such as @Generable and @Guide—let a developer define a Swift type and ask for a typed result. That is generally a better fit than prompting the model to write JSON and then hoping the text parses. Constrained generation improves structural usability, but it does not guarantee that every value is sensible or safe. Validate ranges, required fields, identifiers, and business rules before using the result.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
- 6.9" LTPO Super Retina XDR OLED, 120Hz, HDR10, Dolby Vision, 1320x2868px at 460ppi, 1000 nits (typ), 2000 nits (HBM), 4685mAh Battery
- 1TB, 8GB RAM, Apple A18 Pro (3nm), Hexa-core (2x4.05 GHz + 4x2.42 GHz), Apple GPU 6-core, iOS 18, upgradable to iOS 18.3
- Rear camera: 48MP, f/1.8 (wide) + 12MP, f/2.8 (periscope telephoto) 5x optical zoom + 48MP, f/2.2 (ultrawide), TOF 3D LiDAR scanner (depth), Front Camera: 12MP, f/1.9 (wide)
- 2G: 850/900/1800/1900, 3G: HSDPA 850/900/1700(AWS)/1900/2100, 4G LTE: 1/2/3/4/5/7/8/12/13/14/17/18/19/20/25/26/28/29/30/32/34/38/39/40/41/42/48/53/66/71, 1/2/3/5/7/8/12/14/20/25/26/28/29/30/38/40/41/48/53/66/70/71/75/76/77/78/79/258/260/261 SA/NSA/Sub6/mmWave - Dual eSIM
- Unlocked for freedom to choose your carrier. Compatible with both GSM & CDMA networks. The phone is unlocked to work with all GSM Carriers & CDMA Carriers Including AT&T, T-Mobile, Verizon, Sprint., Etc.
2. Summarize information the app already has
A model can condense a user’s workout history, journal entries, study notes, project updates, or other supplied text. These are strong local-model tasks because the app can provide the relevant source material instead of relying on the model to know current facts. A useful summary should remain grounded in that input; the app should not treat fluent wording as proof that a claim is supported.
3. Personalize explanations and suggestions
Apps can use structured history or user preferences to phrase coaching, create a journaling prompt, explain a learning concept at a chosen level, or describe why a recommendation changed. Keep the authoritative facts and rules in the app. For example, the app should calculate a training metric and determine what recommendations are allowed; the model can help express that result in a clear, personalized way.
4. Add a bounded conversational feature
A study companion grounded in course material, a journaling assistant responding to the user’s own entries, or a game character reacting to game state can feel conversational without promising unrestricted expertise. Make the scope clear to users and provide relevant context to the model. A small on-device model is not a reliable substitute for a general chatbot that can answer open-ended questions about the world.
Rank #3
- 6.1inch Super Retina XDR display. Aluminum with color-infused glass back. Ring/Silent switch
- Dynamic Island. A magical way to interact with iPhone. A16 Bionic chip with 5-core GPU
- Advanced dual-camera system. 48MP Main | Ultra Wide. Super-high-resolution photos (24MP and 48MP). Next-generation portraits with Focus and Depth Control. 4X optical zoom range
- Emergency SOS via satellite. Crash Detection. Roadside Assistance via satellite
- Up to 26 hours video playback. USB C, Supports USB 2. Face ID
5. Let the model request app tools
Tool calling connects language understanding to information or actions the model does not have on its own. An app can expose a tool to search a local catalog, retrieve workout history, perform a calculation, or request current data from a service. The model can select a tool, receive its result, and use that result in its answer. In Apple’s API, tools are defined for the session and their results become part of the interaction.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →This is a practical way to work around a model’s limited knowledge: retrieve current weather, inventory, or a user record from an appropriate source rather than asking the model to guess. But model availability and tool availability are separate. A local database query may work offline; a weather service or remote account lookup may not. Validate tool arguments, enforce permissions, and require user confirmation for consequential actions instead of letting generated intent directly change important data. Apple demonstrates type-safe tool definitions in its WWDC25 session.
How the Swift workflow fits together
The main pieces serve different roles:
SystemLanguageModel.defaultrepresents the available system model and exposes its availability state.LanguageModelSessionis the interaction surface for instructions, prompts, tools, and conversational context.- Ordinary generation is suitable for prose; guided generation is for typed app data.
- Snapshot streaming can expose a partially generated structured value as it fills in, allowing a UI to update progressively.
Start by checking availability and deciding what the app will do if it is absent. A simplified illustrative pattern is:
Rank #4
- This pre-owned product is not Apple certified, but has been professionally inspected, tested and cleaned by Amazon-qualified suppliers.
- There will be no visible cosmetic imperfections when held at an arm’s length.
- This product is eligible for a replacement or refund within 90 days of receipt if you are not satisfied.
- Product may come in generic Box.
import FoundationModels
let model = SystemLanguageModel.default
switch model.availability {
case .available:
// Enable the feature.
case .unavailable(.deviceNotEligible):
// Keep the ordinary non-AI path available.
case .unavailable(.appleIntelligenceNotEnabled):
// Explain the setting; do not block the app.
case .unavailable(.modelNotReady):
// Offer a retry; setup or download may still be in progress.
case .unavailable:
// Provide a general fallback.
}Once the model is available, create a session with instructions that define the task and boundaries, then pass in the user’s request and the minimum context needed. For example, the instruction can tell a workout summarizer not to invent measurements or provide medical advice. Keep app-level rules in instructions and user-supplied material in the prompt or supplied context; neither prompt wording nor model output replaces app-side validation.
let session = LanguageModelSession(
instructions: "Summarize the supplied workout data accurately. Do not invent measurements."
)
let response = try await session.respond(
to: "Summarize this month's training progress: ..."
)The examples are illustrative: check the API against the SDK and OS versions your app supports. Apple’s documentation lists distinct model generations across iOS 26 releases, so API availability and model behavior should be verified for the deployment target. See SystemLanguageModel and the Foundation Models update notes.
For guided generation, define the result as a Swift type rather than asking for JSON in prose:
Best Value
- 6.7inch Super Retina XDR display. ProMotion technology. Always-On display. Titanium with textured matte glass back. Action button
- Dynamic Island. A magical way to interact with iPhone. A17 Pro chip with 6-core GPU
- Pro camera system. 48MP Main | Ultra Wide| Telephoto. Super-high-resolution photos (24MP and 48MP). Next-generation portraits with Focus and Depth Control. Up to 10x optical zoom range
- Emergency SOS via satellite. Crash Detection. Roadside Assistance via satellite
- Up to 29 hours video playback. USB-C, Supports USB 3 for up to 20x faster transfers. Face ID
@Generable
struct WorkoutPlan {
@Guide(description: "The name of the workout")
var title: String
@Guide(description: "A short list of exercises")
var exercises: [Exercise]
}With guided output, the app can work with typed values directly, and snapshot streaming can reveal those values as they become available. This can make it possible to populate sections or cards progressively instead of waiting for a complete answer. Streaming improves the interface’s responsiveness; it does not make a partial result ready to save or act on before validation.
What Apple’s app examples show
Apple has highlighted third-party uses that fit these patterns. In its September 2025 announcement, it presented SmartGym for workout-related generation and coaching, Stoic for journaling prompts, and CellWalk for scientific explanations, alongside VLLO and other apps. At WWDC26, Apple also presented Wayfair and CricHeroes in the context of its newer platform capabilities. These are examples Apple chose to showcase, not independent evaluations or a guarantee that every named feature runs through the same API or model version. See Apple’s announcement and its WWDC26 session.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Choosing between Foundation Models, custom models, App Intents, and cloud AI
| Approach | Best fit | What to keep in mind |
|---|---|---|
| Foundation Models | Use Apple’s system language model for focused, text-oriented features on supported Apple devices. | Availability depends on device eligibility, settings, readiness, and OS support. The model is not intended for open-domain expertise. |
| Core AI | Run a custom model locally when your app needs a model of its own rather than Apple’s system language model. | You own the model choice and its integration, constraints, and evaluation. |
| MLX | Experimentation, research, training, fine-tuning, and local inference workflows, especially on Apple silicon. | It is not simply a turnkey replacement for the system app API or a managed cross-platform cloud service. |
| App Intents | Expose supported app actions and content to system experiences such as Siri and Apple Intelligence. | Complements Foundation Models; it does not itself provide the same in-app generation workflow. |
| Cloud model or backend | Features needing broader knowledge, stronger reasoning, large-scale retrieval, or reach beyond eligible Apple devices. | Network dependence, provider privacy terms, latency, and usage costs need to be designed and disclosed. |
A workout app, for instance, might use Foundation Models to turn a sentence into a draft routine and App Intents to make supported app actions discoverable through Siri. The two solve different problems. Apple’s iOS 26 developer guidance describes App Intents as a way to make app content and actions available to system experiences.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesApple’s WWDC26 materials also discuss a newer provider abstraction that can accommodate Apple and other model providers. Do not assume that provider options shown for a newer SDK are available in every iOS 26 SDK release; verify the documentation for the exact deployment target and API you intend to ship. Likewise, Apple’s announcement of a third-generation model family in June 2026 should not be read as evidence that every iOS 26 device has that newer model. Apple’s announcement covers that later generation.
Where an on-device model can fall short
- Current or open-domain facts: The model is not a dependable source for live information or broad world knowledge. Supply context from an authoritative source or use a different architecture.
- Complex reasoning and high-stakes decisions: Do not make it the sole decision-maker for medical, legal, financial, emergency, or other safety-sensitive outcomes. Use deterministic rules, authoritative services, and appropriate human oversight.
- Large inputs: Long documents and extended conversations may exceed usable context. Break a task into chunks, extract relevant facts, summarize those pieces, and generate a final answer from the compact results. Use token-counting or context APIs where the SDK provides them.
- Unsupported or unready devices: An iOS 26 installation alone is not enough. Keep the core app usable without AI and treat model-not-ready as potentially temporary.
- Network-dependent tools: A local model can be offline while the requested information source is not. Tell users when a feature needs connectivity or cannot retrieve fresh data.
- Changing behavior: System model updates can alter responses. Apple’s update notes call out a model change at iOS 26.4; test against the model versions your app supports rather than depending on a particular phrasing or undocumented behavior.
- Prompt injection in supplied text: Documents, notes, and web content may contain instructions that try to override the task. Treat retrieved content as data, keep tool permissions narrow, and validate every result and action in app code.
Production checklist
- Check availability at runtime. Handle ineligible devices, Apple Intelligence being off, a model still downloading, and other unavailable states. Offer a useful non-AI path rather than blocking the app.
- Design for a specific job. Summarize, classify, extract, transform, or phrase known information. Do not position the local model as an all-purpose expert if the feature cannot support that expectation.
- Constrain and validate outputs. Prefer guided generation for app data, then check numeric ranges, dates, required fields, permissions, and business rules before saving or acting.
- Keep tools least-privileged. Give the model only the tools it needs. Validate arguments, check authorization, and obtain confirmation for actions with meaningful side effects.
- Separate model and data availability. Test offline behavior for both inference and every tool or service the feature calls. Communicate when fresh remote data is unavailable.
- Evaluate across OS/model versions. Keep prompt regression tests for short and long inputs, malformed or adversarial text, and representative edge cases. Test after system model updates and avoid brittle dependencies on exact wording.
- Measure the whole experience. Test latency, responsiveness, and battery impact on supported hardware, including the devices your users actually have. A feature should remain understandable while generation is in progress or unavailable.
- Describe data flows accurately. Explain what stays on device and disclose any analytics, synchronization, or remote tools that transmit data. “On-device AI” alone is not a privacy policy.
For development, Apple says Xcode 26 includes iOS 26 SDKs and requires macOS Sequoia 15.6 or later. A free developer account can be enough to experiment and test on your own devices; paid Apple Developer Program membership is generally relevant when distributing through Apple’s publishing workflow. Check Apple’s current Xcode 26 release notes and developer program page for current requirements and terms.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
