Apple's new Foundation Models framework in iOS 26 lets third‑party apps run a 3‑billion‑parameter on‑device model via the CapGo Capacitor LLM plugin. This eliminates cloud API costs, keeps data on the device, and requires only iPhone 15 Pro/16 or newer, iPad M1 or newer, iOS 26, and about seven gigabytes of free storage for the model weights. The episode walks through initializing the plugin, checking readiness states, handling unsupported devices, and gracefully managing the large background download.
Listeners learn how to work within the 4,096‑token context window by creating task‑specific chats, chunking large inputs, and streaming responses for a smooth UI. The discussion also covers using CapGo Live updates to tweak JavaScript prompts without App Store review, handling built‑in guardrails and refusal errors, and falling back to cloud APIs when on‑device support isn’t available.
Topics:
- initializing the CapGo LLM plugin and checking hardware readiness
- managing the 7 GB model download and graceful UI fallback
- working within the 4,096 token limit with chat segmentation and input chunking
- streaming AI output to improve perceived performance
- using live JavaScript updates to iterate prompts without App Store delays
Full article
https://capgo.app/blog/apple-intelligence-capacitor-app/