A complete AI chat that runs entirely on your Android phone. No account, no cloud, no telemetry. You bring the model file; the app does the rest — offline, private, and yours.
Everything happens locally. Your conversations, characters and files stay on your phone. There is no sign-up, no tracking, and nothing to lose access to. The app makes no network connection at all unless you switch on one of the optional online features — see What talks to the internet below. Both are off by default.
💸 It runs offline so nobody can bill you — which is exactly why nobody bills me. Tip if it earned it:
PayPal — paypal.me/FAMarco
Bitcoin —
bc1qv92c3eyeqvhgfnez7spfd7v2aytkhpshsl65yv
LLM is a self-contained AI chat app for Android. Point it at a model file you already have and start talking — the whole assistant lives on your device. It is built for people who want the power of a modern chat assistant without handing their words to someone else's computer.
No internet connection is required to use it. Airplane mode works perfectly.
- Offline by default. Out of the box the app makes no network connection at all. Your messages are generated on your own hardware.
- Your data stays local. Chats, characters, imported files and settings are stored only in local app storage, encrypted at rest.
- No account, ever. There is nothing to register, nothing to log in to, and no profile to build.
- No telemetry. No analytics, no crash reporting, no phone-home — not even an optional one.
- No lock-in. Your data is yours. Export it whenever you want and take it with you.
Two features can send data off the device. Both are disabled by default and each takes several deliberate steps to switch on. Everything else — the model, your chats, your characters — never touches a network.
- Web search, open webpage, crawl. If you enable them in Settings → Tools and model actions, the model can ask to run a search. Your query text is then sent to the search provider you configured: either your own SearXNG instance, or the Brave Search API. Opening a webpage fetches that page directly. Your chat history and system prompt are never sent — only the query or the URL. Every request is shown to you for approval before it runs.
- Local API server. An optional OpenAI-compatible server for other apps on your own network (Settings → Local API server). It binds to loopback by default; LAN mode requires a generated bearer token. Traffic on your LAN is plain HTTP, so treat it as you would any local development server.
If you never enable either, the app never opens a socket.
- Download the latest release from the Releases page — see the changelog for what changed.
- Allow installation from your browser or file manager if Android asks.
- Open the app, add a model file, and start chatting.
Updating: if you already have v0.14.2 or newer, this installs straight on top and keeps your chats. Coming from anything older than v0.14.2, uninstall once first — the signing key changed in that release.
Every release lists the exact fingerprint of its installer so you can confirm the file is genuine and untampered before installing.
SHA-256: b2cf59e08a0a3d4c59094c3160e222eea1a74666168c22aa1450c758d751ea79
On Windows PowerShell:
Get-FileHash .\LLM-v0.29.0-public-arm64.apk -Algorithm SHA256The value you get must match the one in the release notes exactly. If it does not, do not install the file.
Natural, flowing chat with your own model, with saved history that stays on your device. Pick up any conversation where you left off.
Twenty-one characters are built in and ready the moment you open the app, in two collections.
General holds a senior engineer who answers with the fix rather than a lecture, a stage magician who will actually tell you how the trick is done, a diagnostician who troubleshoots anything broken, a curator who interviews you before recommending a thing, an emergency planner, a wilderness guide, and a night watchman who talks to the museum exhibits. Cyberpunk is the near-future set: privacy and threat modelling, forcing companies to delete your data, street medicine, salvage and repair, biohacking, conditioning, hardware builds, light-up clothing, invention, and security.
Several of them are not conversations but procedures — a typing engine that narrows sixteen personality types down to one, an interrogator who hunts for contradictions in your story, a machine that argues the strongest possible case against your plan, and a text escape room. Each shows its working as it goes.
Tap any card to start a chat where it greets you and stays in role. Import your own cards alongside them, kept in their own section so they never get lost among the built-ins. A set of ready-made behaviour presets lets you shape the assistant from strictly safe to completely unrestricted, and you can write your own.
A built-in RPG mode turns the model into your game master. Pick a setting — or let it surprise you — and play a text adventure with a live character sheet: health, attributes, inventory and status, all kept up to date as the story unfolds. There is no list of options to choose from: you write what you do, in your own words, and the story answers. When what you try could fail, a physical die falls across the screen and decides it — shake the phone while it is in the air and it keeps tumbling until you let it drop. Keep world notes for the places, people and secrets that matter, and the game master stays consistent with them as you go. The whole thing reads like a novel and runs entirely on your phone.
Some of the strongest models available are far bigger than any phone's memory — and this app runs them anyway, reading the parts it needs from storage as it goes. A 24 GB model on a phone that has nowhere near 24 GB free is an ordinary Tuesday here.
Every model tells you how much free memory it needs before you tap it, and the drawer shows how much your phone actually has, live. If something is loaded and you need the memory back, one button hands it over.
When a reply is long, the app keeps it going on its own instead of stopping halfway — you get the whole answer without prompting for more.
On newer devices the app uses on-device acceleration to answer faster, and picks the right setting for each model automatically — no tuning required. Whether acceleration is available at all depends on the model file you bring; see Which model file should I get? below.
A clean, modern interface with light and dark modes, adjustable text and spacing, high-contrast and reduced-motion options, and a range of accent colours to make it yours. Answers render properly — including tables — and a built-in Help section explains the options in plain language.
Turn your phone into a private AI endpoint on your home network, so your other apps and devices can use your on-device assistant — still without touching the cloud.
Optional app lock with PIN or fingerprint, encrypted backups, and a private-screen mode keep your conversations for your eyes only.
Your phone has two processors that can run a model, and which one your model gets to use is decided largely by the file you download — so it is worth thirty seconds before you pick one.
The two processors, in plain terms:
| What it is | What it's good at | |
|---|---|---|
| CPU | The general-purpose processor every phone has | Always works, with any model. The reliable default, and on most phones the fastest one too. |
| GPU | The graphics chip | Sometimes helps, sometimes doesn't — the app measures it once and then keeps the winner. |
The one thing you control: the quantization.
Model files come in "quantizations" — compression levels, written in the filename like Q4_0 or Q4_K_M. This is not only a quality dial; it decides whether the GPU can run the file at all, because the graphics chip physically understands only certain formats.
Q4_0— the safe bet. The format the GPU path accepts. If you want to give the GPU a chance, get this one.Q4_K_M,Q5_K_M,Q6_K— the most common downloads, marginally better quality, but CPU-only. If your model never leaves the CPU, this is almost always why.
Do not over-think it: the CPU path is well optimised (ARM KleidiAI, wide instruction support) and on many devices simply wins. A Q4_K_M model on the CPU is a perfectly good default.
Curious what your model is actually using? Settings → Device and performance names the exact reason for the model you have selected.
- A modern Android phone with enough memory and storage for the model file you intend to run.
- A compatible model file (which you provide) — see Which model file should I get? if you want on-device acceleration.
- No internet connection required.
Proprietary. All rights reserved. This repository contains the app's public documentation and releases only; the source code is not open.
© eVersor-HN