SCX.ai launches Australian-hosted AI for Apple apps

SCX.ai is building an Australian-hosted AI model integration for Apple's Foundation Models framework. The project would make it the first Australian group to offer that option to developers using Apple's software tools.

The integration, called SCXSwiftKit, is a Swift package that connects SCX.ai's hosted text models to Apple's framework for AI features in iPhone, iPad and Mac apps. It is designed to let developers choose between Apple's on-device AI and models hosted by SCX.ai while continuing to use Apple's existing development interface.

The move reflects a broader shift in software design as AI spreads from standalone chatbots into everyday applications. Developers are increasingly deciding whether a task should stay on a device or be sent to a remote model in a data centre, a choice that can affect data residency, privacy and computing cost.

SCXSwiftKit conforms to Apple's LanguageModel interface and is intended for distribution through Swift Package Manager. Developers can add the package, select an SCX.ai model for a session, and continue working with Apple's LanguageModelSession, @Generable, Tool and streaming application programming interfaces.

That means one app could use Apple's local model for a simple task and an SCX.ai-hosted model for another that needs more computing resources or a larger context window. Developers would not need to build around a separate AI framework.

Local hosting

A central part of the offering is Project MAGPiE, which SCX describes as its Australian sovereign model. Hosted in Australia, it offers a 131,000-token context window aimed at work involving longer documents.

Available models are resolved at runtime, so newly added text models can be made available without requiring developers to update the package. SCX's wider catalogue also includes models hosted in the United States, Germany and Singapore. Each model exposes available regions as country codes so applications can check where data will be sent before a prompt is submitted.

For businesses and software developers, that regional visibility comes as data handling becomes a more prominent issue in AI deployment. A consumer may not know whether an AI feature is running directly on a handset or whether the request is being processed in another jurisdiction.

"Apple is making AI part of the basic plumbing of software development," said David Keane, Founder and Chief Executive Officer of SCX.ai.

"When millions of developers can start building AI directly into the apps people use every day, the amount of inference required is going to grow enormously.

"For most people, they won't necessarily know whether the AI they are using is running on their phone or whether their request is being sent somewhere else to be processed. That makes the choices developers make about where AI runs increasingly important.

"We want Australian developers to have a genuine Australian option, where they can choose Australian-hosted AI when data residency matters."

Inference demand

SCX is positioning the project around a sharp increase in inference demand, the computing needed every time a user makes an AI request. That differs from the model training phase, which has drawn much of the investment attention in recent years.

The company runs SambaNova inference processors from its first node at the Equinix SY5 data centre in Sydney. Its Australian infrastructure is intended to increase the amount of inference delivered from the available power supply and physical data centre space, rather than focus on dedicated training sites.

Keane linked that approach to the economics of everyday AI use as more software products embed language models and AI agents.

"Everyone is talking about the enormous computing required to train AI models," Keane said.

"The next challenge is what happens when AI is being used all day, every day, by hundreds of millions of people. The economic question becomes: how much useful AI can you produce from every unit of electricity?"

SCXSwiftKit supports text responses and streaming, tool calling, @Generable structured output, and Light, Moderate and Deep reasoning levels. Apple has said developers can use third-party models through the same session interface used elsewhere in its Foundation Models framework, giving outside providers a route into Apple's broader software ecosystem.

The project also extends a long-standing connection between Keane and Apple. His previous companies, BigTinCan and Zunos, were original members of Apple's Mobility Partner Program in 2015 and among the first iPad software developers in that initiative.

Keane said the industry is now entering another platform shift.

"Not every AI task can be handled on a phone and not every AI task needs to be sent to a giant data centre," he said.

"The future is hybrid. Some AI will happen on the device and some will require more powerful infrastructure. For consumers, the important question will increasingly be: when my AI leaves my device, where does it go?"