Qwen, Llama, SmolLM, Phi and more. MobileLLM rates each one for your phone before you download, then runs it on the phone itself. No PC, no server, no account.
Ratings follow the app’s own rules for the Q4_K_M file of each model.
Everything happens in two screens: Model Management to choose and download, Chat to talk.
Open a model and each GGUF file is listed by quantization with its size and a rating worked out from your RAM and storage. Files that would crawl, like IQ quantizations or 3B-class models, say why before you spend the data.
Answers stream in as they are written. Every bubble shows how long it took and how much of the context is used, the conversation is saved on the phone, and older turns are trimmed to fit when a chat gets long.
Ask for an alarm, a call or a message and the model can prepare it. Nothing runs until you see exactly what will happen and tap Run.
The model runs on your phone’s CPU. Once a model is downloaded, MobileLLM needs no internet at all, and there is no account, server or analytics on our side.
We run no server and ship no advertising, analytics or crash-reporting SDK. There is nowhere for your chats to go.
The model list and files come straight from Hugging Face. What you type is never part of those requests.
Loading a model and chatting with it work with no connection at all.
Chats and models live in the app’s own storage. Delete a model, clear the chat or remove the app and they are gone.
Small models like Qwen2.5 0.5B and Llama 3.2 1B run on phones with about 4 GB of RAM. The app sizes its context to your memory: 1,024 to 4,096 tokens.
On start, MobileLLM picks the fastest build your processor supports, from ARMv8.0 up to ARMv8.6 with i8mm matrix instructions.
A dropped connection or a closed app does not start a 2 GB download over. It picks up where it stopped, and warns you first on mobile data.
The list starts with SmolLM2 1.7B, Qwen2.5 0.5B and 1.5B, Llama 3.2 1B and 3B, Gemma 2 2B and Phi-3.5 mini, each with every GGUF file its Hugging Face repository offers.
Start with a Q4_K_M file of a 0.5B to 1.5B model. Larger files and IQ quantizations are slower on a phone CPU, and the rating says so before you download.
These are small models. They are quick to try and fun to compare, but they make mistakes and can invent facts, so check anything important.
No. MobileLLM runs on the CPU, using the fastest instructions your processor has. That behaves the same on every Android phone.
The app is in English and Korean. Models answer in the language you write in, as well as their training allows.