Apple brings on-device AI to third-party apps

Apple just made on-device AI completely free for developers, and the App Store is already changing.

Following the rollout of iOS 26, iPadOS 26, and macOS 26 this month, Apple officially detailed how developers are tapping into its Foundation Models framework. The core promise is straightforward: full access to the local large language model powering Apple Intelligence, running on-device, offline, with zero API inference costs.

Fitness and journaling go fully local

A slew of popular apps have already pushed updates to show off what the local model can do.

SmartGym now lets users describe workouts in plain language, turning rough prompts into structured routines complete with sets, reps, and equipment swaps. The app's Smart Trainer explains the exact reasoning behind weight changes and generates dynamic greetings based on fitness stats. Matt Abras, CEO of SmartGym, called it a game-changer that made previously impossible on-device features simple to build.

Stoic is using the local model to power hyper-personalized journaling. If a user logs poor sleep or a bad mood, the app serves up supportive prompts and notifications based on recent context. Founder Maciej Lobodzinski noted that features once demanding heavy back-end infrastructure now run natively on-device, keeping sensitive personal thoughts strictly private.

Other sports tools are chiming in as well. SwingVision combines Core ML video tracking with the foundation model to deliver actionable tennis and pickleball advice. Train Fitness suggests alternative exercises when gym gear is occupied, while 7 Minute Workout dynamically adjusts routines around nagging injuries.

Education and task managers jump in

The local model is also finding a home in classrooms and task lists.

In CellWalk, an interactive 3D biology app, users can tap complex scientific terms to trigger conversational explanations. Developer Tim Davison used tool calling to ground responses in the app's verified data, tailoring explanations directly to the user's knowledge level.

Stuff developer Austin Blake used the framework to ship both Listen Mode and Scan Mode at once. Users can speak disjointed thoughts or snap photos of messy handwritten notes, and the app instantly parses dates, tags, and lists entirely on-device.

Grammo adds real-time grammar quizzes, Lil Artist pairs the model with the ImageCreator API for children's stories, and Vocabulary automatically sorts saved words into thematic buckets like verbs or anatomy.

Why it matters

For developers, zero-cost inference changes the unit economics of mobile apps overnight. Small studios and indie hackers no longer have to fear cloud token bills every time a user asks a question, removing the biggest barrier to adding AI features.

For users, it means everyday apps gain genuine intelligence without handing over intimate health or diary entries to third-party servers. Plus, everything keeps working smoothly when offline.

We think Apple's play here is remarkably clever. Instead of trying to outmuscle cloud giants in brute-force reasoning, Apple is leveraging its hardware footprint to make local AI the default utility across its platforms. The lingering catch, however, is that compact on-device models still hit walls with complex, multi-step logic. But for everyday summaries, natural language parsing, and quick prompts, Apple just gave developers an offer they can't afford to pass up.