Crashes for me on a couple of different iDevices (2 generations behind) after only a few 2-3 chats. Probably not enough RAM.

Saw this one on X the other day updated with Gemma 4 and they have the built-in Apple Foundation model, Qwen3.5, and other models:

Locally AI - https://locallyai.app/

I'm able to sweet talk the gemma-4-e2b-it model in an iphone 15 to solve a hcaptcha screenshot. This small model is surprisingly very capable!

Gemma 4 is great: https://aibenchy.com/compare/google-gemma-4-31b-it-medium/go...

I assume it is the 26B A4B one, if it runs locally?

It's so ridiculous that Google made a custom SoC for their phones, touting its AI performance, even calling it Tensor, and Apple is still faster at running Google's own model.

Google really ought to shut down their phone chip team. Literally every chip from them has been a disappointment. As much as I hate to say it, sticking with Qualcomm would have been the right choice.

This is fun. I just wish I could add more skills, the UX is too dumbed down but knowing there is a run_js tool there is a lot that can be done here.

E4B is pretty good for extracting tables of items from receipt scans and inferring categories, wish this could be called from within a shortcut to just select a photo and add the extracted table to the clipboard

Extremely impressed by how fast responses are on iPhone 17 Pro Max. Can’t wait for this to be used for Siri’s brain one of these days (hopefully!)

I need normal server too in local network I can run chat in other device and 'counting' on iphone.

Second idea is input audio in other language, like Czech, Polish, French

Still didnt release training recipe, data, methodology etc unlike deepseek. Mostly released to get developer ecosystem across their android built in ai. Still good and interesting, but not exactly philanthropic to the open source progress.