In Abhinav Raj’s November 8, 2026 account for XDA, an iPhone reached a model running on a PC through LM Studio, Locally, and LM Link. The PC performed the inference; the iPhone acted as a client. Raj reported connecting over cellular with the phone’s Wi-Fi off, without opening a router port.
How the PC-hosted setup works
A host is the computer that runs the AI model; a client sends it requests and displays its responses. In Raj’s setup, LM Studio v0.4.28 ran Google’s 4-bit QAT build of Gemma 4 26B A4B on a PC with an RTX 4070 Ti Super. Locally on the iPhone connected to that computer through LM Link. The phone did not run that model.
Raj reported that the model’s weights alone occupied about 14.4 GB. In his configuration, he kept 30 layers on the GPU and used a 16K context window. Those are details of his particular setup, not a general hardware requirement.
LM Studio lists Locally for iPhone and iPad. Raj’s account describes signing in to Locally, connecting through LM Link, then selecting the PC and model. For broader context on where mobile AI processing happens, see NeoTeo’s Spanish-language explainer on mobile AI privacy.
With remote access, the PC must be available to do the work. Running inference on the phone is a different setup with different constraints.
Cellular access still depends on the PC
Raj said he turned off Wi-Fi on his iPhone, switched to cellular data, and still reached the PC-hosted model without opening a router port. That describes his test, not a guarantee for every network or configuration.
The computer remains a necessary part of the arrangement. Raj reported that it had to stay on and awake; when Windows put it to sleep, the phone lost its connection to the hosted model. If you want to use this route away from the desk, the PC must remain reachable and awake.
A separate route runs the model on the iPhone
Better Stack’s presenter Warren demonstrated Qwen3.5-35B-A3B running directly on an iPhone 17. He reported 11 tokens per second in thinking mode and said the phone got hot during the run. Those observations describe that demonstration, not a benchmark against Raj’s PC-hosted setup.
The demonstration used Flash iOS, based on the FlashMOE engine. Warren described keeping about 1.4 GB of model weights in memory while streaming other model components from storage. The setup shown involved Xcode and a paid Apple developer account. This is a separate on-device route, with its own setup and thermal considerations.
Choosing between the two approaches
Remote access lets the iPhone use a model hosted on a PC, but the computer must stay awake and reachable. Direct on-device inference runs the model on the iPhone itself; Warren’s demonstration reported heat during use. The examples involve different models and configurations, so their reported speed does not establish which approach is faster for a given task.
