Hey, Siri: Plug into my app like you said you'd do way back in 2024

As of Tuesday, developers can explore expanded Siri AI integration for their Apple platform apps through beta releases of iOS 27.3, iPadOS 27.3, macOS 27.3, and visionOS 27.3. Next spring, nearly three years after announcing plans for deeper third-party Siri integration, Apple expects to begin accepting App Store submissions using the new capabilities.

"This developer beta offers a sneak peek at Siri AI app intents for developers," Apple said in a developer news post. "With this early Developer Preview, your app can work directly with Siri AI, giving you a first look at new ways to reach and serve your customers."

Anything is likely to be an improvement. Apple's Siri, a voice-activated assistant service, dates back to 2011 and gained limited third-party app access in 2016. In 2024, it became an interface for interacting with Apple Intelligence, its suite of generative AI features.

Apple Intelligence has not been very intelligent - several advertised Siri capabilities were delayed, so much so that Apple was sued in 2025 over allegations of misleading marketing. After a shakeup of Apple's machine learning group in December 2025, Siri and Apple Intelligence received a makeover at Apple's Worldwide Developers Conference in June.

"Before this year, Siri was a timer with a voice," said Roberto Gutierrez, Apple platforms staff engineer at a media company and author of the e-book On-Device Intelligence, in an email to The Register. "Weather, alarms, calls. Anything past that was a coin toss.

"On the developer side, SiriKit covered a handful of domains and Shortcuts only worked if users memorized the phrase. Siri AI is the first version that understands context and can chain actions across apps. That's a real change."

There is a catch, however, Gutierrez observed. Siri AI is available in a limited number of languages and regions. Support for Traditional Chinese is expected by the end of 2026, and EU availability remains a matter of negotiation - Siri AI is unavailable on iOS, iPadOS, and watchOS in the bloc as Apple works through interoperability concerns under Europe's Digital Markets Act.

"Apple is working constructively with the EU in an attempt to find a path forward," the company said last month without committing to any concessions.

Gutierrez said that he expects opening Siri to third-party developers will be useful to some extent.

"Siri can only do what apps expose through App Intents, and SiriKit is on its way out," he said. "If your app's main actions aren't exposed, the new Siri can't see you. The bigger shift is letting Claude, Gemini and others plug in. Developers already choose between three options for every AI feature: Apple's system model, an open model through MLX, or a cloud API. Siri is heading the same way. It doesn't need to be the smartest model. It needs to be the best router. That's the right call."

Apple has made choosing between those three options more difficult by failing to publish benchmarks that show how its Foundation Models compare with open-weight alternatives.

The biz has published measurements of how its third-generation Foundation Models compare to their 2025 predecessors. But the company's commitment - to "[share] more details on our latest foundation models, including updated evaluations and benchmarks, in a technical report later this summer" - remains unfulfilled so far as we're aware.

Gutierrez ran his own benchmarks to compare Apple's iOS system model (FM system, ~3B) to several open-weight models that are suitable for use on-device.

"On an A17 Pro iPhone running iOS 26, I tested Apple's system model against Qwen3 1.7B and 4B, Llama 3.2 3B and Gemma 3 4B on a structured-output task," he said. "Speed: Apple's model ran at about 149 tok/s, roughly 3x the fastest open model (Qwen3 1.7B, about 46)."

Quality, he said, was a tie, with every model passing five typed checks. But memory, he said, was where Apple's models had an advantage.

"The system model's weights live outside your app's memory budget," he explained. "My app peaked at 12 MB during inference. An open 4B model costs 2.4-2.7 GB, and a 2.4 GB model plus your app plus a WebView is a crash report. So yes, it's competitive, and on iPhone it's the default I'd recommend. If you want your own model, Qwen3 1.7B is the one."

But Gutierrez urged Apple to publish proper competitive benchmarks. "Their numbers compare against their own previous models," he said. "Developers aren't choosing between Apple 2025 and Apple 2026. We're choosing between Apple, Qwen and a cloud API. I had to build my own harness to answer a question Apple could settle with one table."

The absence of competitive model metrics is not the only gap. Apple has also failed to accommodate web apps.

"Web apps are locked out," said Gutierrez. "Foundation Models is a Swift API (and now Python). No JavaScript, no WebKit. A PWA's only option is to ship its own model into the browser and pay the full memory bill inside a WebView."

And bringing your own model may pose performance problems. "On an A17 Pro iPhone, the 3-4B open models pushed the phone into 'fair' thermal state and dropped to 12-24 tok/s," he said. "In practice, your own model on iPhone means ~2B or smaller."

With Siri AI unavailable on iPhones, iPads, and Apple Watches in the EU (and its rollout in mainland China subject to regulatory requirements), developers may need a fallback for global apps to handle regional restrictions. Apple's Private Cloud Compute offers eligible developers access to Foundation Models, Gutierrez observed, though it requires membership in the App Store Small Business Program, fewer than two million first-time App Store downloads across their apps, and the necessary entitlement. (R)