Running a language model on your phone used to feel like a novelty, and now there’s a whole aisle of apps that actually make it practical. Reading the store listings, you’d struggle to tell any of them apart though - they all look identical and you don’t actually know how each one holds up after you’ve spent some time with it.

The only way to know which one to choose is to run them as your only mobile AI for an extended period. I started experimenting with mobile local AI earlier this year, and at this point I’ve tried about six runners. However, only a couple of them are really worth the install in my opinion, and even then, only one actually checks all my boxes. For reference, I’m running these on my iPhone 16, which has 8GB of unified memory.

The easiest way to get on-device AI

Google AI Edge Gallery is the app I’d point a total newcomer to, and I mean that as a compliment. Google builds it and keeps it open source, and it’s a proper App Store download now rather than the sideloaded file it started as. The engine is Google’s LiteRT, built specifically for phone chips, and my battery hardly reacts to running this app for extended periods - at least, way less than streaming or playing a game.

There’s quite a lot of control tucked away inside the tidy interface surface. It has most of the configurations a desktop runner has, including temp, toK, topP, context length, CPU/GPU acceleration, and speculative decoding. But my favorite part is still the Agent Skills. This opens you into a chat window with options for using an interactive map, setting reminders, tracking your mood, querying Wikipedia, and generating QR codes. This is why Edge Gallery is the best local alternative to Google Assistant, because many of their functions overlap.

The model selection is where it may disappoint. Edge Gallery only accepts LiteRT-format models, and that catalog is tiny beside the GGUF world every other app draws from, so you wind up with Gemma whether you meant to or not. Importing a LiteRT build from Hugging Face is possible, but almost nothing I actually wanted came in that format unless I’m willing to convert it myself.

The iOS build also has more rough spots; the Mobile Actions do close to nothing on an iPhone, so using this app to its full capacity is best suited for Android users. But the chat is still solid for iOS.

PocketPal

The option for tinkerers

PocketPal has been my mobile default longer than anything else on this list, and unlike the others it runs on Android as well as iOS, so most people reading this can use it on whichever phone they want. It runs on llama.cpp and instead of a fixed shelf you pull models straight from Hugging Face - there are hundreds of them in GGUF. Compared to Edge Gallery, that’s a whole different world, and it’s going to be your route past the Gemma family.

I run Qwen 3.5 2B (Q4_0) and Qwen 3.5 4B (Q3_K_M) on here; they’re much better suited for anything research or learning-adjacent. It’ll also run more niche models like H2O’s Danube or a Phi build, so you can really match the model to the job.

The control is the real perk. You get temperature, context length, and also penalty parameters, as well as a built-in benchmark that shows what your hardware can take before you commit to a download. Pals let you save personas and presets, and newer builds even added in-chat web search if you bring an API key. It’s really the full package, and this is what I’d recommend to those who’ve already been running local LLMs on PC.

The one thing it won’t do is read a document. There’s no way to attach a file or hand it a PDF, so you have to paste the text in yourself. For pure chat, that doesn’t really matter, and it held its spot as my main app because pure chat was most of what my phone got used for.

Noema

The obscure app running formats the big names can’t

Noema is the app I see barely anyone in this corner talking about, and I have a feeling that might be because it’s iOS-only - local LLM users seem to lean more Android due to its hardware flexibility. So if you have an iPhone, this is the app I couldn’t recommend enough.

It runs GGUF like the rest, but also Apple’s MLX and Liquid AI’s own SLM format, which is what will get you past the usual Qwen and Gemma builds if you want to experiment more. An MLX model tuned for Apple silicon i

Note: The source article appears to have been cut off at this point.