Issue 719

25^th July 2025

Written by Dave Verwer

Comment

Thanks so much to everyone who filled last week’s survey about Apple’s on-device Foundation Model. I’ll start this with a quick summary of your responses:

1. The @Generable macro works well for small data structures but less so for larger ones.
2. It’s good at tasks that don’t require knowledge, like summarisation, sentiment analysis, and data categorisation.
3. The 4096 token context is limiting, and (predictably) results get worse when reaching the context limits.
4. The safety refusals/warnings are too easily triggered in surprising situations.
5. Many people who experimented have switched back to using a hosted LLM, as iOS 26 won’t be available on enough devices until next year.

However, I also wanted to experiment with the model myself. I started writing a little sample app until I found the Foundation Models Framework Example repository by Rudrank Riyam. What a demo app! I’m confident there’s no better demonstration and playground for the Foundation Model, and I’m incredibly grateful for his work on it.

I experimented with some prompts and confirmed the model could summarise well. Then, I was curious how it would guide the output into something usable in an app by adhering to a word and paragraph limit and to write in first person. I’ve tried this before with an earlier iteration of ChatGPT, and it struggled with word counts. I was also curious how it compared to other LLMs, both locally through LM Studio and the online ones you’ve all heard of.

So, I did a little experiment. I imagined a hypothetical journaling app, grabbed a random entry from a famous journal and wrote a prompt for it to create an index-style view in an app. Then, I picked a couple of models to run locally that approximately matched the Foundation Model (~4k context and a few billion parameters) and ran the prompt past all of them.

I expected Apple’s model to be comparable to the other local models, and for them all to lag behind the online models. Check the results for yourself, but I was impressed with the Foundation Model. It followed my instructions perfectly and performed measurably better than the other local models. It even competed well (for this task) with the online ones. For the right kind of task, it appears to be a real alternative to all those expensive API calls.

If you take away anything from this, it should be to download and play with the Foundation Models Framework Example app I mentioned earlier and find small tasks it could do inside your apps.

For example, I tested another idea by asking the model to tell me whether a piece of text was either a “document title” or the “start of a paragraph”. Imagine if Apple Notes processed your first few characters of typing into a new note and removed the default “Heading” style if it thought you were typing a paragraph. 🤯 The Foundation Model was reliable at getting the answer to that question correct.

What could it do for your apps?

– Dave Verwer

Sponsored Link

Stream’s iOS SDK lets you add production-ready chat, voice/video calling, activity feeds, and AI moderation to your app with just a few lines of code. Backed by a global edge network and a generous free Maker plan, you can prototype today and scale to millions. Start Coding Free.

News

New age rating categories are coming with this year’s operating systems, and you’ll need to re-classify all your current apps by January 31st, 2026. That’s not super urgent yet, but why not get it done now?

It’s worth remembering that the iOS 26 and iPadOS 26 public betas started this week. Luckily, beta testers can no longer leave App Store reviews, but that doesn’t mean you won’t get a few emails to your support department if you’ve anything in your apps that will break on the betas. Thanks to Dan Moren for the reminder that Apple pushed the public beta button.

Tools

This tool from Sahil Satralkar looks interesting. Its primary task is to find unused images in your app, but it also has a whole host of other smaller utilities like de-interlacing PNG files and checking that all your images are in the same colour space. I’ve not tried it, but it’s the kind of thing that might make a good final check before submission.

Code

You might think you know the answer to the question John Sundell asks in this blog post title. You probably do know most of it, but I bet there’s at least one nugget of information in here that’s new to you. Who knew you could write ~2,000 words on this subject? 😱

I always love reading a good debugging story based on a real world issue. So, how do you think I felt about Pol Piella Abadia’s latest post which has three! They are all from the Helm app, and include a crash, a performance regression, and an unexpected system prompt.

Design

David Smith was surprised how little needed to change as he moved his Pedometer++ watchOS app from the watchOS 10 design standard to Liquid Glass. However, that doesn’t mean the post isn’t interesting. The changes are subtle, but I love how he walks us through his design process.

Books

This new course from Victor Baro looks great. Its title might make you think it was going to focus on the SwiftUI side of things, but before you know it you’ll be diving into coordinate spaces, colour maths, and advanced animation using Metal directly. The course is free, but there’s also an optional paid component with bonus effects, examples, and a full Xcode project showing it all in practice.

For full disclosure, Victor sent me a review copy of the Xcode project.

Jobs

Senior Software Engineer, iOS @ Doximity – Doximity, the medical network used by over 80% of US clinicians is seeking a talented Sr iOS Software Engineer to join our remote-first team! If you’re an expert in Swift with a strong understanding of TCA, and eager to make a real impact on healthcare, we want to hear from you. – Remote (within US timezones)

Sr iOS Developer @ Hero Assistant – We seek to build the most ambitious iOS app ever created. This is a once-in-a-lifetime opportunity to join a rocket ship early, backed by world-class investors, and already proven by tens of thousands of users who rely on Hero every day. – On-site (United States in CA)

iOS/macOS Developer @ LIT Software – LIT Software is a small, highly specialized team that develops powerful litigation tools for Apple platforms. Our industry-leading apps help lawyers organize, analyze, and present evidence with ease. Join us in building the best tools for legal professionals. – Remote (within US timezones)

And finally...

Sorted. 🧮