A small on-device model can’t be great at everything — but it can become the right specialist on demand. This post ships one frozen base model plus a library of small LoRA adapters, then loads, swi...

Step-by-step: from a 1-bit .gguf model to a working SwiftUI app generating text on a real iPhone — with the bridge code, the build, and the gotchas that actually cost me hours (a reasoning model th...

Step-by-step: from a .litertlm model file to a working SwiftUI app generating text on a real iPhone — with every layer, command, and gotcha. 💻 Full source code: the complete sample app for this...

A measured look at where the GPU wins, where it loses, and the one ratio that decides it — running gemma-3-270m on an iPhone XS. TL;DR I deployed a 270M-parameter LLM on an iPhone XS and benchm...

After three months of hands-on research into on-device AI, I’ve come to realize this field is not simply about knowing how to code models — it sits at the intersection of machine learning, systems ...

2025 marked a critical threshold: for the first time, on-device AI inference hit sub-20ms latency on mid-range Android phones for production computer vision models. Not on a server. Not on a $1,200...