Ask a few iOS developers what “building an AI app” means right now and you will get a few different answers. Some mean wiring up Apple’s on-device Foundation Models framework. Others mean calling out to a cloud model provider over a REST API. Most end up doing a bit of both.
iOS 27 did not settle that question so much as it made the answer more important to get right. Apple Intelligence now sits right alongside your own backend, and figuring out where each one belongs is quickly becoming a core design decision rather than an afterthought.
This post walks through how AI apps on iOS 27 actually get built in practice, where on-device models stop being enough, and why REST APIs are still doing most of the heavy lifting once your app needs to go beyond what ships in the box.
What iOS 27 gives you on-device
With iOS 27, Apple expanded the Foundation Models framework, giving developers direct access to the same on-device large language model that powers Siri AI and the system’s own Apple Intelligence features. This matters because it runs locally, works offline, and does not require you to manage API keys or pay per token. For tasks like summarizing text, extracting structured data from a document, or generating short pieces of writing inside your app, the on-device model is often genuinely good enough.
The catch is scope. The on-device model is smaller and more constrained than frontier cloud models, and Apple intentionally keeps it focused on narrower tasks rather than open-ended reasoning or long, complex generations. It is also only available on Apple Intelligence-compatible hardware, which still leaves a meaningful slice of the install base without access. For a lot of AI apps on iOS 27, on-device models cover the easy half of the feature list and leave the harder half for something else.
Where REST APIs take over
That harder half is where REST APIs come in, and for most AI apps, this is still the backbone of the real functionality. Anything involving longer context, more complex reasoning, multimodal input, or access to a model with capabilities beyond what ships on-device means calling out to a provider like the Claude API or another cloud model over REST. The same is true for anything that depends on live data your app does not have locally, like pulling fresh information from a database or a third party service before the model generates a response.
This creates a fairly common architecture for AI apps on iOS 27: use the on-device Foundation Models framework for fast, private, offline-capable tasks, and fall back to a REST API call for anything that needs more horsepower or external data. Getting that split right, rather than routing everything through one path out of habit, is quickly becoming one of the more interesting design decisions in iOS development.
Why testing the REST side matters more, not less
It is tempting to assume that leaning on Apple’s on-device model takes pressure off your own backend, but the opposite tends to be true. Once part of your AI feature runs locally and part runs through a REST API, users expect both paths to feel equally fast and equally reliable. A REST API call that is slow or inconsistent stands out more, not less, when it sits next to an instant on-device response for a similar kind of task.
This is especially true if your app also integrates with Siri AI through App Intents, since those intents often depend on the same REST API calls your in-app AI features use. A flaky endpoint does not just cost you a bad in-app experience anymore. It can break a Siri AI interaction too. Testing your REST API thoroughly, including authentication flows, streaming responses, and how it behaves under real network conditions rather than a fast office Wi-Fi connection, matters more now that it sits in more places at once.
Testing AI APIs without slowing down development
Switching between Xcode, a terminal running curl commands, and whatever client you use to inspect a cloud provider’s API response adds friction to a workflow that is already fairly complex. This is where having a native API client built for this exact situation pays off. HTTPBot gives you a native REST API app on iPhone, iPad, and Mac that handles REST requests, WebSocket connections, and GraphQL queries, with support for streaming responses and authentication through API keys, OAuth 2.0, or JWT, which covers how most AI providers gate access to their models.
For developers building AI apps on iOS 27, that means you can test a request to your model provider, check how a streaming response comes through, and confirm your JSON parsing handles the payload correctly, all without writing a throwaway script first. Since HTTPBot syncs collections across your Apple devices through iCloud, you can build and test a request on your Mac, then pick the same collection back up on your iPhone to confirm it behaves the same way on a cellular connection, not just your desk’s Wi-Fi.
Choosing the right split for your app
There is no single correct ratio between on-device and REST-based AI in an app, and that is probably a good thing. A note-taking app might lean heavily on-device for quick summarization and only reach out to a REST API for more complex rewrites. A customer support app might do the opposite, using on-device models for lightweight triage and a REST API for anything that needs real context from a live backend.
What matters is making that choice deliberately, based on what each task actually needs, rather than defaulting to whichever path was easier to wire up first.
Wrapping up
Building AI apps on iOS 27 is not a choice between on-device models and REST APIs. It is a question of which one fits each part of your feature set, and increasingly, most serious apps need both working well together. The on-device side gets a lot of the attention right now, but the REST APIs behind your cloud model calls are still doing the heavier lifting, and they deserve the same testing rigor as everything else in your stack.
If you want a faster way to build and verify that side of things, HTTPBot brings native REST, WebSocket, and GraphQL testing to your iPhone, iPad, and Mac, so your AI features hold up wherever they are running.
