What Are Apple Foundation Models? Apple’s On-Device AI, Explained

“Apple Intelligence” is the brand. Apple Foundation Models are the actual language models underneath it, and since 2025 Apple has let any app use them through the Foundation Models framework. Here is what they are, what they can and can’t do, what changed in iOS 27, and how to try them yourself without writing a line of code.

Two models: on your device and in Apple’s cloud

Apple’s system is built around two kinds of model:

1. The on-device model. A small language model that runs on the chip in your iPhone, iPad or Mac. Apple describes it as roughly 3 billion parameters, compressed to about 2 bits per parameter so it fits in memory and runs fast without draining the battery. Because it runs locally, it works offline and your data doesn’t leave the device.
2. The server model on Private Cloud Compute. A larger model that runs on Apple-silicon servers for heavier requests. Apple says data sent there is used only for your request, never stored, and not accessible to Apple, and that independent experts can verify those promises.

For comparison, the models behind ChatGPT, Claude and Gemini are far larger and run only in data centers. Apple’s bet is that a small model on your device, backed by a private cloud for harder jobs, covers most everyday tasks.

What the Foundation Models framework lets apps do

The framework is a native Swift API. With it, an app can:

• Generate text from a prompt, with instructions that set the model’s role and tone
• Stream a response as it’s produced
• Hold a session, a multi-turn conversation where the model remembers earlier turns
• Use guided generation, where the model fills in a Swift type the app defines, so the output is structured data rather than loose text
• Call tools, functions the app provides, such as searching the web or reading a calendar, and then answer using the results

All of this runs on the device, with no API key and no per-request bill.

What changed in iOS 27

At WWDC 2026 Apple expanded the framework considerably. According to Apple:

• Images: multimodal prompts let apps send images alongside text, and on-device Vision tools let the model reason about them.
• Dynamic Profiles: apps can swap models, tools and instructions within one continuous session.
• Evaluations: a new framework for testing that AI features behave reliably.
• More models: the same API now reaches Apple’s models on Private Cloud Compute and any provider whose Swift package conforms to Apple’s Language Model protocol.
• Free cloud for small developers: apps in the App Store Small Business Program with fewer than 2 million first-time downloads can use the next generation of Apple’s Private Cloud Compute models at no cloud API cost.

Limits worth knowing

• Small context. Apple’s developer documentation gives the on-device model a 4,096-token context window per session, shared by instructions, prompts, tool definitions and responses. That is a few pages of text. Since iOS 26.4, apps can read the current size from the system, because Apple can change it as it updates the model.
• Smaller model, smaller knowledge. A 3-billion-parameter model knows far less than frontier models and is weaker at long reasoning. It is built for summarizing, rewriting, extracting and short answers, not deep research.
• Its own guardrails. The model declines some requests under Apple’s rules.
• Device and language limits. It runs only on Apple Intelligence devices, in the languages Apple supports.

Why it matters: private and nearly zero-energy AI

On-device models change two things about AI.

Privacy. When the model runs on your phone, your prompt never has to leave it. There’s no server log of what you asked.

Energy. Every cloud AI answer is computed in a data center that draws electricity and often evaporates water for cooling. An on-device answer uses only a sliver of your battery. There is no data center at all. Read more in is on-device AI more private and eco-friendly?

How to try Apple’s on-device model yourself

You don’t need Xcode. Viro AI is built on the Foundation Models framework and lists Apple Intelligence in its model menu next to ChatGPT, Claude, Gemini and Viro’s own model.

• On an eligible iPhone, iPad or Mac, choose Apple Intelligence and chat. The answer comes from your device.
• In Viro it can also search the web, read your calendar and files, and ask your connected apps, using the framework’s tool calling. Only that one request leaves your device.
• It’s free on every plan, and Viro counts each reply as 0 Wh because no data center is involved.
• When you need more, switch to a larger model in the same chat.

Step-by-step: how to chat with Apple Intelligence like ChatGPT.

Try Apple Intelligence in Viro AI, free on iPhone, iPad and Mac.

Frequently asked questions

What are Apple Foundation Models?

They are the language models Apple built for Apple Intelligence: a small model that runs on your iPhone, iPad or Mac, and a larger model that runs on Apple’s Private Cloud Compute servers. Apps can use them through Apple’s Foundation Models framework.

How big is Apple’s on-device model?

Apple describes its on-device foundation model as roughly 3 billion parameters, compressed to about 2 bits per parameter so it fits and runs efficiently on the device.

What is the context window of Apple’s on-device model?

Apple’s developer documentation gives a context window of 4,096 tokens per session, shared by instructions, prompts, tool definitions and responses. Since iOS 26.4, apps can read the current size from the system, because it can change as Apple updates the model.

Is the Foundation Models framework free?

The on-device model runs on the user’s device, so there are no per-request API charges. Apple also says apps in the App Store Small Business Program with fewer than 2 million first-time downloads can use its Private Cloud Compute models at no cloud API cost.

What’s new in Foundation Models in iOS 27?

Apple says the framework now accepts images alongside text, can call on-device Vision tools, lets apps swap models, tools and instructions mid-session with Dynamic Profiles, adds an Evaluations framework, and works with Apple’s Private Cloud Compute models and other providers that conform to its Language Model protocol.

How can I try Apple’s on-device model without coding?

Use an app built on the framework. Viro AI lists Apple Intelligence as a chat model, so on an eligible iPhone, iPad or Mac you can chat with Apple’s on-device model directly, for free.

Sources

1. Apple Intelligence for developers, Apple Developer, checked November 2026
2. Apple Intelligence Foundation Language Models Tech Report 2025, Apple Machine Learning Research, 2025
3. TN3193: Managing the on-device foundation model’s context window, Apple Developer
4. Apple Intelligence, Apple, checked November 2026
5. Apple Intelligence in Viro, Viro Help Center

Cover photo by Brecht Corbeel on Unsplash