(Credit: Apple)
CUPERTINO, Calif.—As we speak, software developers around the world are exploring the inner workings of Apple's foundational AI models for the first time. They got access through a developer beta on June 9, a major though little-discussed announcement at WWDC 2025.
Opening up the 3-billion-parameter model through the new Foundation Model framework is a "huge step," says Craig Federighi, Apple's SVP of Software Engineering. "We think this will ignite a whole new wave of intelligent experiences in the apps users rely on every day. We can't wait to see what developers create."
Anyone in the Apple Developer Program can try it out. Full access to all tools requires a $99-per-year membership, though that might be cheaper than paying to access models from other brands, depending on the need. A public beta will be available next month.
The models are also only available when programming apps using Apple's Swift coding language. Plus, Apple Intelligence is limited to Apple's newest devices, so it might be a while until we see what these models can really do. Still, my hope is that giving developers access will clarify where Apple's technology stands, and how it compares to rivals like OpenAI and Google.
Will the models offer enough capabilities to entice developers to use it? If so, will the app features relying on them work as expected?
(Credit: Emily Forlini/PCMag)To Date, We've Questioned Apple's Intelligence
I've found the details of Apple's AI opaque since it first debuted at WWDC 2024. I noticed that Apple did not name the models, like OpenAI does with GPT-4o, or Google with Gemini 2.5, for example. It's just "Apple Intelligence," a phrase the company uses for all-things AI; consumer-facing features, backend technology, branding, you name it. That left me wondering whether Apple actually trained and developed something new, or if this so-called intelligence is just a big exercise in rearranging the furniture.
So far, the features powered by Apple Intelligence have been hit or miss. Several are half-baked, like last fall's Siri updates, and the Image Playground, which creates odd, animated images of people in your Contacts. Apple might be aware that it's a little weird; at WWDC, it announced that it's incorporating ChatGPT's image generator into the Image Playground.
In fairness, some Apple Intelligence-backed features announced at WWDC are impressive, but they follow those already available on rival devices. That includes Live Translation (Google has this on Meet) and Apple's new "Visual Intelligence" feature, a concept Google pioneered with Circle-to-Search. Apple hasn't yet demonstrated that it has the chops to lead the way in AI.
Why Apple's Models Might Be Worth Using
Opening the models could give us a better sense of what Apple is capable of, though some devs are struggling to muster excitement for Apple Intelligence, The Wall Street Journal reports, particularly when they could experiment with larger models from OpenAI or even DeepSeek.
Even if the models end up being limited, which we don't know yet to be true, there are a few reasons they could be worth using. For one, there's the fact that Apple customized the models to work seamlessly on its devices. Easy access may also encourage developers to at least experiment with them for basic features. The models also support 15 languages and work offline.
Apple also offers the strictest privacy promise in the AI biz. By keeping the data "on-device," the tech doesn't harvest data for third parties or send it to far-flung servers, aside from Apple's Private Cloud Compute network. That's well-suited to education apps like Kahoot!, which could use the on-device model to "generate a personalized quiz" from a student's notes. Apple also cited AllTrails, which could use AI to let hikers find trails suited to their needs while off-the-grid.
The data comes from publisher licensing agreements, open-sourced datasets, and "ethical webcrawling" with the Applebot. Apple says it does not use "users' private personal data or user interactions when training our foundation models," and it removes "certain categories of personally identifiable information and to exclude profanity and unsafe material."
"The on-device model is optimized for efficiency and tailored for Apple silicon, enabling low-latency inference with minimal resource usage, while the server model is designed to deliver high accuracy and scalability for more complex tasks," Apple says. "Together, they form a complementary suite of solutions adaptable to diverse applications."
Maybe we should all cross our fingers that developers are satisfied with the models, especially if the on-device privacy promise holds up. That's a bit more warm and fuzzy than using China's DeepSeek, which is known for being cheap and powerful—and also sending data to the CCP. But it remains to be seen if Apple's models perform well and actually manage data responsibly. The next year should be telling, for Apple users and the company's AI ambitions.
For more on Apple's AI strategy, check out this take from our sibling site, CNET: Apple's de-chatbot-ification of AI is nearly complete.
If you're a developer who's experimenting with Apple Intelligence foundation models, please reach out to share your initial impressions (emily.forlini@ziffmedia.com).