Run language models on your phone. Offline, private, no cloud.
An Android app powered by llama.cpp, with GGUF models from
Hugging Face, that also talks to your LM Studio, Ollama and llama-server.
Download · Features · What's new in this fork · Models · Build · Contribute
Personal fork of andriydruk/LMPlayground, with extra features layered on top of the original app. Full credit for the original goes to Andriy Druk; like upstream, this fork is MIT-licensed.
Get the latest APK from Releases, then open it on your phone and allow installs from your browser or file manager.
| File | Use it for |
|---|---|
local-llm-<version>-arm64-v8a-sideload.apk |
Daily use: smaller and smoother (R8-shrunk, not debuggable). |
local-llm-<version>-arm64-v8a-debug.apk |
Fallback, and the one to use when reporting a crash. |
- Needs an arm64 phone on Android 11 or newer.
- It installs alongside the Play Store LM Playground and does not touch its data.
- This fork is not published to the Play Store. Every release also lists
SHA256SUMS.txt. - No release yet for your commit? Every push builds an APK on GitHub Actions: open a green run and download the artifact.
- On-device inference: no cloud, no API keys, fully offline
- Rich markdown in chat responses: headers, code blocks, lists, and more
- Reasoning model support: thinking steps from models like GPT-OSS, DeepSeek R1, and Nemotron are shown in a styled section
- Tool calling: capable models can search the web and Wikipedia, fetch a page, run JavaScript in an on-device sandbox, and use a calculator, unit converter, date and time, device info, location and a memory
- Image input: attach a photo to models with vision (Qwen 3.5, Gemma 3 4B, Gemma 4 E2B/E4B, Ministral 3, LFM2.5-VL); the small image add-on downloads the first time you need it
- GPU and speculative decoding: OpenCL GPU backend on arm64, and an optional, experimental MTP speculative-decoding switch for Qwen 3.5 and Gemma 4
- Reliable background downloads: custom download engine with OkHttp and WorkManager, progress notifications with speed and ETA, automatic resume on network interruptions
- Storage management: choose where to keep multi-GB model files with Android's Storage Access Framework
- ARM optimized: KleidiAI kernels and OpenMP for faster generation on arm64 devices
- Large-screen ready: tablets, foldables, and Chromebooks get a permanent sessions sidebar, list-detail Settings, and freeform window resize support
Features added on top of upstream LM Playground:
- Message actions: copy any message, edit & resend one of your prompts, or regenerate the latest reply (LM Studio-style). Editing or regenerating rebuilds the conversation from that point and re-runs generation.
- Remote servers (LM Studio, Ollama & llama.cpp): save as many OpenAI-compatible servers as you like (name, address, optional API key), typed by hand or found by scanning the network. When you open the model picker, each saved server is checked and only the ones that are online appear, each as its own entry where the available models are grouped by provider (Qwen, Gemma, …) and listed by their cleaned-up names. Remote chat streams with reasoning, accurate tokens/second, and a server-side offload when you unload. A model-details card (frosted, matching the picker) shows the server model's real quantization, context window, and live generation-parameter sliders. Away from home, point it at the server's Tailscale IP for the same thing LM Studio's "LM Link" does.
- Project folders: organize chats into collapsible folders in the sessions drawer; deleting a folder also deletes the chats inside it.
- Redesigned reasoning panel: the model's "thinking" and tool-call details render as collapsible rounded cards with an animated chevron. While the model reasons the card stays live with a running token count and a rotating status label (e.g. "Thinking…", localized in English and Italian), and the thinking text itself is rendered as markdown.
- Generation stats: an optional line under each reply showing total tokens, elapsed time, tokens/second, and time-to-first-token. Toggle it in Settings → Sound, Haptics & Stats.
- Context-window meter: a small circular gauge next to the composer shows how full the model's context window is, from the engine's real KV-cache usage, with a permanent percentage beside it (LM Studio-style). The arc shifts color as the window fills; tap the ring to toggle the exact used / total token count.
- Green rebrand: a distinct green app icon and the name "Local LLM" to tell this fork apart from the original at a glance.
- One-tap builds: a GitHub Actions workflow builds an installable APK on every push, and tagging a version publishes it as a release with notes taken from the in-app changelog.
See the changelog for the full version history.
| Family | Sizes | Provider |
|---|---|---|
| GPT-OSS | 20B | OpenAI |
| Qwen 3.5 | 0.8B, 2B, 4B | Alibaba |
| Qwen 3 | 0.6B, 1.7B, 4B | Alibaba |
| Gemma 4 | E2B, E4B, 12B | |
| Gemma 3n | E2B, E4B | |
| Gemma 3 | 1B, 4B | |
| Nemotron 3 Nano | 4B | NVIDIA |
| Granite 4.2 | 3B | IBM |
| Granite 4.1 | 3B, 8B | IBM |
| Granite 4.0 | Micro, H-Tiny | IBM |
| DeepSeek R1 Distill | 1.5B, 7B | DeepSeek |
| Phi-4 mini | 3.8B | Microsoft |
| LFM2.5 | 350M, 1.2B (Instruct & Thinking), 2.6B | Liquid AI |
| LFM2.5-VL | 450M, 1.6B, 3B | Liquid AI |
| SmolLM3 | 3B | Hugging Face |
| MiniCPM5 | 1B, 2B | OpenBMB |
| LFM3 | 1.7B | Liquid AI |
| Spark-X2.6 | 1.7B, 4B | XHToken |
| Spark-X2.5 | 1.7B, 4B | XHToken |
| Ministral 3 | 3B, 8B (Instruct & Reasoning) | Mistral |
| Llama 3.2 | 1B, 3B | Meta |
Legacy models
| Family | Sizes | Provider |
|---|---|---|
| Qwen 2.5 | 0.5B, 1.5B | Alibaba |
| Phi 3.5 mini | 3.8B | Microsoft |
| Mistral v0.3 | 7B | Mistral |
| Llama 3.1 | 8B | Meta |
| Gemma 2 | 9B |
Most models use Q4_K_M quantization (Qwen 3.5 too: the GPU backend has kernels for Q4_K but none for Q3_K). Gemma 4 uses Google's official QAT Q4_0 builds, and GPT-OSS ships in its native MXFP4 format. See ModelInfoProvider.kt for the full list.
Prerequisites:
- Android Studio Otter (2025.2.1) or newer (the project uses Android Gradle plugin 8.13) with JDK 21, which Android Studio bundles
- NDK 27.2.12479018
- CMake 3.31.5
- Clone the repository with submodules:
git clone --recurse-submodules https://github.com/Leo00703/Local-LLM.git
- Apply the Android patches to the llama.cpp submodule. CI does this on every build; a local checkout needs it once, and the native code does not compile without it (details in the patches README):
cd app/src/main/cpp/llama.cpp
for p in ../patches/llama.cpp/*.patch; do git apply "$p"; done
cd -
- Open the project in Android Studio:
File>Open> Select the cloned repository. - Connect an Android device or start an emulator.
- Run the application using
Run>Run 'app'or the play button in Android Studio.
Or build the APK from the command line (after step 2):
./gradlew assembleDebug # output in app/build/outputs/apk/debug/
No local toolchain? Push to any branch (or run the Build Debug APK workflow from the Actions tab) and download the resulting APK artifact.
Bug reports, ideas and pull requests are welcome: start with CONTRIBUTING.md, and use the
issue templates. AGENTS.md has the architecture
map and the build/CI notes. Security problems go through SECURITY.md, not a public issue.
This project is licensed under the MIT License, inherited from upstream LM Playground.
This is a fork of LM Playground by Andriy Druk, built on llama.cpp. Models are GGUF-format (mostly Q4_K_M quantization) sourced from Hugging Face.