🚀 just dropped the first-ever gemma 3n model running fully on ios — yes, on-device, no cloud. ￼

📱 it’s slow (for now), but it benchmarks nearly on par with claude 3.7 sonnet for non-coding tasks. ￼

🧠 gemma 3n is google’s new mobile-first frontier model, designed for edge devices. ￼

🔧 i built a simple ios app to run it locally:

💬 features: on-device inference, real-time chat ui, streaming responses.

⚠️ it’s a bit slow, but it’s a start.

👀 try it out and see the future of on-device ai.

🚀 just dropped the first-ever gemma 3n model running fully on ios — yes, on-device, no cloud. ￼

📱 it’s slow (for now), but it benchmarks nearly on par with claude 3.7 sonnet for non-coding tasks. ￼

🧠 gemma 3n is google’s new mobile-first frontier model, designed for edge devices. ￼

🔧 i built a simple ios app to run it locally:

💬 features: on-device inference, real-time chat ui, streaming responses.

⚠️ it’s a bit slow, but it’s a start.

👀 try it out and see the future of on-device ai.

https://github.com/sid9102/gemma3n-ios

i dabble with ai using open identity data from crypto https://alexpaden.tech

Barista in the morning, AI engineer by day, and parent (with @shiwen) by night ||
Host at /clickcast /cpg

@alexpaden @jachian you two will love this. Gemma3n running locally on an iOS device

Farcaster Development Boutique @ https://dTech.vision 💜 in love with tech 🤓decentralization, fitness and oldschool rap 😎

Sadly the google implementation only supports CPU acceleration, which is why it's so slow. Hoping for them to release open weights in a non-proprietary format soon!

awesome thing is it's pretty good at following instructions. I've got it reliably outputting JSON, trying to use it for more than just text gen. tool use on-device is possible now!

It’s true, but the fact that we can put Gemma on a phone tells you it’s progressing. For now specialty LoRa models are an interesting design space to play in