Apple's new Foundation Models framework gives developers direct access to its ~3B parameter on-device language model. Features deep Swift integration for easy use, powerful privacy, and cost-free AI inference for apps on Apple silicon.
Hi everyone!
I know a lot of the talk after WWDC has been about the new Liquid Glass design, but I think the long-term impact of Apple's progress with their on-device models is being underestimated.
With the new Foundation Models framework, Apple is giving developers direct access to the ~3B on-device model that powers many Apple Intelligence features. The integration with Swift makes it remarkably simple to bring generative features into an app, all while running locally, privately, and at no inference cost.
Given Apple's massive hardware install base, this framework could unleash a huge amount of potential. While on-device AI still has a way to go to catch up with the most advanced cloud models, I personally believe this gap will close faster than many expect. Once on-device AI crosses that key capability threshold, we're likely to see a whole new wave of app innovation.
some evals straight from Apple's dev docs for Foundation Models framework:
- US english: on par with Llama 4 scout and Qwen 3 235b
- non-US: beats Llama 4 scout
- non-english: worse than qwen
generally much worse than old gpt 4o
and when you have to confine benching to the failure that is llama 4, and its smallest model at that
ya this is why you don't talk about Siri rofl 😭
Everyone’s discussing the new liquid glass and the UI overhaul and I think this new framework just doesn’t get all the attention it deserves. If done right, this can be the crucial missing “AI feature” every Apple fan and investor was expecting - instead of competing with others’ AI features, Apple tries to provide the right toolset for the developers to build those. Apple was always the best in providing powerful, easy-to-use frameworks and SDKs and the App Store was always the biggest advantage of iOS and iPadOS. I think Apple does the right thing by betting on it once more.
This sounds like a major leap forward for app development! How flexible is the Foundation Models framework when it comes to fine-tuning the on-device model for specific app use cases?
