I'm still a little dissapointed. It seems those models are only available for iPhone series 16 and iPhone 15 pro. According to mixpanel that's only 25% of all iOS devices and even less if taking into account iPadOS. You will still have to use some other online model if you want to cover all iOS 26 users because I doubt apple will approve your app if it will only work on those Apple Intelligence devices.
Why should I bother then as a 3rd party developer? Sure nice not having a cost for API for 25% of users but still those models are very small and equivalent of qwen2.5 4B or so and their online models supposed equivalent of llama scout. Those models are already very cheap online so why bother having more complicated code base then? Maybe in 2 years once more iOS users replace their phones but I'm unlikely to use this for developing iOS in the next year.
This would be more interesting if all iOS 26 devices at least had access to their server models.
Uptake of iPhone 16+ devices will be much more than 25% by the time someone develops the next killer app using these tools, which will no doubt spur sales anyway.
App development could be as quickly as a few weeks. If the only "killer apps" we have seen in the past three years are the ChatGPT kind, I'm not holding my breath for a brand new "killer app" that runs only on iPhone 16+.
It's not a killer until it's verified to work, and makes a better product than those that use APIs. From a user's perspective, nobody cares about whether it's on-device or not unless cost is directly involved.
The user doesn’t necessarily care but I think on-device enables certain experiences that can’t be done if you involve a cloud service.
For example, involving a network and data transfer will always toss in some latency that may not be acceptable or desired for a particular use case, even if you’re on a desktop plugged in with an Ethernet cable
I also think that in the long term the hardware that individuals own is far more powerful than the typical compute that can be allocated to a user who is using a free or cheap cloud service. Comparing a rented VPS to a cheap $400 server in my closet is like night and day, the VPS is just not a lot of horsepower and the typical smartphone has a whole lot of computing power to work with. In a very near future when the chips are even more AI-optimized, data centers might not be the most efficient way to go about this.
Example: imagine implementing an application like Final Cut Pro as a browser application where all the compute takes place on a remote cloud server. It’s just not plausible: too much data to handle, too much compute needed for processing, too much of a need for low-latency responsiveness of the app.
Why should I bother then as a 3rd party developer? Sure nice not having a cost for API for 25% of users but still those models are very small and equivalent of qwen2.5 4B or so and their online models supposed equivalent of llama scout. Those models are already very cheap online so why bother having more complicated code base then? Maybe in 2 years once more iOS users replace their phones but I'm unlikely to use this for developing iOS in the next year.
This would be more interesting if all iOS 26 devices at least had access to their server models.