Consequently, I likely won't be able to utilize Apple's native Siri services. Because of this, the cloud-based Siri model currently remains a gray area for MiniSwift.

But it’s not a major setback. I’ve overcome plenty of hurdles before, and I’m confident I’ll find a way around this one too ;-)

After extensive trial and error, I built my own foundation model and successfully embedded it into this project, leveraging WebGPU once again.

Since Apple’s 3B foundation model wasn't highly efficient to begin with, I was confident that I could replicate and reflect the core capabilities it offered.

However, with the upcoming Apple 27 update, it looks like they are transitioning to a completely cloud-based architecture.

As for the AI side of things, I’d like to share a few thoughts.

Initially, the Foundation library that came with macOS 26 was more ideal for my needs, as it featured a 3B local model. This aligned perfectly with MiniSwift’s core philosophy of offline-first operation.

So, how did I tackle the challenge when things shifted?

I developed an end-to-end LLM architecture called https://llm.istanbul written entirely in WebGPU, giving me absolute control over the system.

@mrudokas @joethephish I wasn't expecting an apology anyway ;-) We followed each other in a very civilized way. Unfortunately, this Claude issue keeps coming up, but it's a theme I've worked hard to establish not just here, but on all my pages: llm.istanbul, toprak.run, sengi.surf, bla bla.

Thank you.