PocketLFM

Run Liquid AI's LFM2.5 large language model fully on-device on Android — offline, private, no cloud.

_{on-device LLM · edge AI · LFM2.5 · Liquid Foundation Model · offline AI Android · llama.cpp Android · local LLM · private AI · GGUF · edge inference}

---

🏆 What makes this notable

To our knowledge, PocketLFM is the first independent open-source Android app that runs Liquid AI's LFM2.5 fully on-device via llama.cpp — outside Liquid AI's own Apollo app and LEAP SDK.

Liquid AI claims LFM2.5 is built for the edge and runs under 1 GB of memory. This app is an open, llama.cpp-based reference implementation that demonstrates exactly that — no proprietary SDK, just GGUF + native inference on your phone.

---

💡 What is PocketLFM?

PocketLFM is a native Android chat app that runs a real LLM — Liquid AI's LFM2.5-1.2B-Instruct — entirely on your device. No internet. No servers. No data ever leaves your phone.

It uses llama.cpp's C++ inference engine (via the Android NDK) with a modern Jetpack Compose UI. Because it speaks the GGUF format, it can also run any small GGUF model — TinyLlama, Phi, Llama 3.2, Gemma, etc. — but it ships pointed at LFM2.5, which is purpose-built for edge devices.

🔒 100% offline	Inference runs locally. Zero network calls for chat.
🧠 LFM2.5-1.2B	Liquid AI's edge-optimized model, ~750 MB (Q4_K_M), <1 GB RAM
⚡ Native speed	C++ / NDK backend via llama.cpp
🎨 Modern UI	Jetpack Compose, MVVM, Hilt, Coroutines/Flow
🔁 Any GGUF	Swap in any GGUF model with one URL change
🛡️ Private by design	Chat history stays in the app's private storage

---

📲 Try it

Pre-built APK coming soon. For now, build from source (below) — it's a standard Gradle/Android Studio project.

The app downloads the LFM2.5 GGUF on first launch (~750 MB, one time), then runs fully offline.

---

🔧 Build & Run

Prerequisites

• Android Studio Hedgehog (2023.1.1) or newer
• Android SDK 24+ (Android 7.0+)
• NDK 25.0+ and CMake 3.22+ (install via SDK Manager)
• A 64-bit ARM device with 4 GB+ RAM (LFM2.5 runs comfortably under 1 GB)

Debug build

git clone https://github.com/Jeevav62/pocketlfm.git
cd pocketlfm/lfm25
./gradlew installDebug

From Android Studio

1. Open the lfm25/ folder in Android Studio
2. Sync Gradle
3. Pick a device/emulator (API 24+, ARM64)
4. Run ▶️

On first launch the app fetches the LFM2.5 model. After that it's fully offline.

---

🔄 Using a different model

PocketLFM works with any GGUF model. Change one line in ModelDownloader.kt:

const val DEFAULT_MODEL_URL = "https://huggingface.co/.../your-model.Q4_K_M.gguf"

Then update the chat template in ChatViewModel.kt (buildPrompt) to match the model's expected format. LFM2.5 uses ChatML (<|im_start|> / <|im_end|>).

---

🧱 Tech stack

Component	Technology
Language	Kotlin
UI	Jetpack Compose
Inference	llama.cpp (C++ / NDK)
Model	Liquid LFM2.5-1.2B-Instruct (GGUF)
Build	Gradle + CMake
Architecture	MVVM
DI	Hilt
Async	Coroutines & Flow

---

📦 Release APK

# 1. Create a keystore (first time only)
keytool -genkey -v -keystore release.keystore -alias pocketlfm \
  -keyalg RSA -keysize 2048 -validity 10000

# 2. Configure signing in local.properties (never commit this):
#    store.file=release.keystore
#    store.password=...
#    key.alias=pocketlfm
#    key.password=...

# 3. Build
./gradlew assembleRelease     # APK  -> app/build/outputs/apk/release/
./gradlew bundleRelease       # AAB  -> app/build/outputs/bundle/release/

⚠️ Never commit keystores or passwords. local.properties is git-ignored. Use Google Play App Signing for distribution.

---

🗺️ Roadmap

• Pre-built signed APK in Releases
• On-device benchmarks (tokens/sec across devices)
• Model picker (switch models in-app)
• LFM2.5-VL (vision) support
• Voice in/out (STT/TTS)
• RAG over local documents

---

🙋 Work With Me

I'm Jeeva — I turn research papers into shipping products: reading the latest work and rebuilding it into real features, tools, and ideas people can use. PocketLFM is one example — taking Liquid AI's edge-AI claim and proving it with an open, on-device implementation anyone can run.

What I do:

• 🔬 Research → product — take papers/models and ship them as usable features
• 🗣️ Fine-tune TTS & LLMs — voice cloning, custom speech models, domain-tuned language models
• 🛠️ Practical ML & audio tooling — pipelines, datasets, and apps that actually run

I'm open to collaboration, open-source work, and hiring opportunities. If you need someone who reads the paper and ships the product, let's talk:

• 🤗 Hugging Face: huggingface.co/jeevav62
• 🌐 Portfolio: see my work

---

📜 License

MIT — see LICENSE.

🙏 Acknowledgments

• Liquid AI for the LFM2.5 models that make edge LLMs practical
• llama.cpp by Georgi Gerganov for the inference engine
• The Android Jetpack Compose team