Day 12/14 • can local AI replace Gemini?

Apple just introduced a brand new API that most developers still don't know about. It's called the Attachment API, and for the first time ever, it allows developers to access native, multimodal AI capabilities. This means developers can finally analyze text and images 100% on-device and fully offline. I just spent 8 hours fully pushing the capabilities of this model to the limit.

I was curious if it could replace my current image processing stack, which is a mix of Gemini models, since this new model is trained on Gemini. So I tried a lot of different experiments in my bookshelf scanning app. Quick pro tip, don't test in your simulator since it won't be accurate. Instead, I had Claude build an AI lab onto my phone to run the test and automatically send the results back to my laptop.

Unfortunately, I was never able to get it above 40% accuracy, which is way below my current 98%. But that doesn't mean it's bad. I tested simpler things like describing objects in an image, and it was super fast and accurate. But most importantly, 100% free and offline, whereas I'm spending up to 10 cents on every scan with Gemini, and I require a decent internet connection. So I can definitely see a lot of useful applications.

I'm Brian, and this is day 12 of building my first app in 14 days. Follow along for the next part where I finally submit to the App Store.

Follow along for the next part where I finally submit to the App Store.