Takeaways Regarding AI at WWDC 26
What's Changed with On-Device AI on Apple Platforms
Hello everyone,
Apple released some amazing new features for AI and Apple Foundation Models this week at WWDC, and while we didn’t get exactly what I was hoping for, we got some amazing new features. Here’s the highlights that I think were important.
What We didn’t Get
We did not see an increase to Apple Foundation Models’ context window. It looks like all on device models still have a limit of 4096 tokens. This is unfortunate as a larger context would have allowed for more chat experiences. The second negative is that there is no longer adapter support for the system model. This may be negated though if we can fine tune the core model and use it as a Core AI model. Let’s now turn to the positive aspects now that those negative aspects are out of the way, because it is completely good news from here on.
New Models
Apple did announce the names of their new models. They include AFM 3 Core, AFM 3 Core Advanced, AFM 3 Cloud, AFM 3 Cloud Pro, and AFM 3 Cloud Image Specialized. You can look up the details of these models yourself, but they have some amazing features, and we can take advantage of these models in our apps. The AFM 3 Core Advanced model is quite interesting as we’ve learned that it is what allows for the new voices in Siri AI, and the new speech to text features that were announced. This model requires devices that have 12 GB RAM or more.
Apple Foundation Models With Other Models
Another change that was made this year is that AFM can load MLX models and now, Core AI models. My understanding is that Core AI models are similar to CoreML, but are designed to run regular models that can perform chat and other functions. Examples include Qwen 3 and Gemma 3. All of these models can now run straight from Apple Foundation Models, which I think is incredible! Another neat thing is that AFM can load models from cloud providers as well, and you can also take advantage of Private Cloud Compute as long as your developer account doesn’t have more than 2 million downloads.
Final Thoughts
I was really hoping to see an increase in context for the built in models, but it really does look like Apple has made some big changes in how on device AI works on Apple hardware. It is amazing to think that models like Gemma 4 can run on the same playing field as the built in models if we use Core AI or MLX. We can just use the same APIs now to get the job done. All of these changes will come out in September, and many things will change between now and then, so I am very excited to try all of this and see what’s possible in my apps, and with the platform as a whole.