LLM Hub 🤖

LLM Hub is an open-source mobile app for on-device LLM chat, image generation, video generation, and music generation, available for both Android and iOS, and now on Windows, macOS and Linux as LLM Hub Desktop. It's optimized for mobile usage (CPU/GPU/NPU acceleration) and supports multiple model formats so you can run powerful models locally and privately.

🌟 Featured: Featured by RunAnywhere on their website — read the LLM Hub iOS Case Study!

Download

💻 LLM Hub Desktop

LLM Hub Desktop 0.1.2 is out: native apps for Windows, macOS and Linux. Chat with local models, run Claude Code and OpenClaw on them, turn prompts into slides and pages, generate images, video and music, and serve OpenAI and Anthropic compatible APIs to your other apps. No account and no cloud. See the release notes for what's new and checksums.

Download	For
macOS	Mac with Apple silicon, macOS 26 or later. Signed and notarized.
Windows x64	Windows 10 (version 2004) or later on an Intel or AMD PC
Windows on Arm	Snapdragon PCs
Linux .deb	Ubuntu 24.04 or later, Debian and their derivatives
Linux AppImage	Other x64 distributions with GTK 4 and libadwaita 1.4 or later

The Windows installers aren't code signed yet, so Windows may show "Windows protected your PC": click More info, then Run anyway. To build it yourself, see desktop/README.md.

📸 Demo & Screenshots

Vibe Coder Demo on iOS	Image Generation on Android
Vibe Coder using Gemma 4 model on iPhone (HTML preview)	Stable Diffusion image generation on Android

AI Agent Termux Command Execution on Android	Music Generation on iOS
Autonomous AI Agent executing Termux shell commands (ls list files) with live terminal output on Android	On-device music generation with Magenta Realtime 2 on iPhone

🚀 Features

🛠️ AI Tools Suite

Tool	Description
💬 Chat	Multi-turn conversations with RAG memory, web search, TTS auto-readout, and multimodal input
🤖 AI Agent	[NEW] Autonomous device agent with function calling, interactive maps, device tools, MCP tools, and Termux terminal command generation on Android
🤖 creAItor	[NEW] Design custom AI personas with specialized system prompts (PCTF) in seconds
💻 Vibe Coder	[NEW] Build apps with live HTML/JS preview; Android also integrates Codex through Termux using your selected local model
✍️ Writing Aid	Summarize, expand, rewrite, improve grammar, or generate code from descriptions
🎨 Image Generator	Create images from text prompts using Stable Diffusion 1.5 with swipeable gallery
🎵 Music Generator	[NEW] Generate music and sound effects locally with SoundGen on Android and Magenta Realtime 2 on iOS, including live playback as music is generated
🔎 Instant Media Search	[NEW] Search photos and videos by description or a reference image on Android and iOS using on-device EmbeddingGemma 2
🎞️ Video Moment Finder	[NEW] Describe a moment in a video and jump to matching timestamps using visual and audio embeddings on Android and iOS
🔊 Text to Speech & Voice Cloning	Android supports Kokoro, multilingual Supertonic 3 with voice-profile import, and Pocket TTS with on-device voice cloning
🔍 Image Upscale	[NEW] Upscale images up to 4× using AI super-resolution models (RealESRGAN, UltraSharp) with NPU acceleration
🎥 Video Generator	[NEW] Generate videos from text prompts or images: LTX-Video on Android (Snapdragon NPU, Vulkan GPU or CPU), Stable Video Diffusion on iOS
🌍 Translator	Translate text, images (OCR), and audio across 50+ languages - offline
🎙️ Transcriber	Convert speech to text with on-device processing using Whisper models

🤖 AI Agent & Termux Terminal Command Generation

The AI Agent turns on-device LLMs into autonomous assistants with device function calling capabilities:

• Device Actions & Function Calling: Embedded interactive maps (OpenStreetMap), calendar management, alarm setup, flashlight toggle, SMS & Email composition, cryptographic hashing (SHA-256/MD5), and math calculation.
• MCP Tools: Connect Model Context Protocol (MCP) servers so the Agent can discover and use external tools such as filesystem, search, document, memory, or custom hosted tools. MCP calls require user approval before execution.
• Termux Command Generation (Android): Automatically drafts shell/CLI commands targeting Termux environment execution (RUN_COMMAND_SERVICE intent interface).
• Interactive Pre-Execution Editing: Users can inspect and edit AI-generated shell commands before execution.
• Live Dark Terminal Container: Formats and streams live command outputs (stdout/stderr) in a dark monospace terminal UI block.
• AI Error Auto-Correction Loop: When a shell command fails or encounters permission/syntax errors, the agent automatically feeds the error logs back into the LLM for self-diagnostics and auto-generates fix commands.

🎙️ Supported ASR Models (Android)

The offline Transcriber feature on Android supports on-device Whisper models for high-accuracy local speech-to-text.

🔊 Android Text to Speech & Voice Cloning

• Kokoro: Local speech with downloadable voices managed in Settings.
• Supertonic 3: Multilingual on-device speech with downloadable presets and import of compatible voice-style JSON files. Raw audio cloning and Pocket TTS voice states are not supported by Supertonic's importer. See Supertonic 3 setup.
• Pocket TTS: English speech and fully local voice cloning from recorded or imported audio. Manage, select, rename and delete cloned voices in Settings. Cloning uses 3–30 seconds of reference audio; longer imports use their first 30 seconds. See Pocket TTS setup.
• Selecting an engine and voice in Settings applies it to app-wide speech and auto-readout. Translator keeps system TTS.

🔎 Media Search & Video Moment Finder

On Android and iOS, Instant Media Search uses EmbeddingGemma 2 to index photos and videos locally, then find matches from a text description or reference image. Video Moment Finder combines video frames and audio from two-second windows so you can search for a scene or event and play the matching moment. Download a compatible EmbeddingGemma 2 model and choose the media to index; search runs on-device.

💻 Android Vibe Coder & Codex

Enable Codex mode in Vibe Coder's model settings to use the selected on-device model with the Codex app server running in Termux. Codex can inspect project files, edit code and run project commands, with streamed activity and terminal output in the app. This mode requires an ARM64 device running Android 10 or newer and does not require an OpenAI login. Install Termux, allow external app commands, grant LLM Hub permission to run Termux commands, and use the in-app Install Codex action before starting.

🎥 Android Video Generation

The Android Video Generator runs Lightricks LTX-Video 2B 0.9.8 distilled fully on-device, from a text prompt or a starting image, in 8 denoising steps.

• Snapdragon NPU: The transformer and VAE decoder run on the Hexagon NPU as QNN context binaries, one build per Hexagon generation: v69 (8 Gen 1, 8+ Gen 1), v73 (8 Gen 2), v75 (8 Gen 3), v79 (8 Elite) and v81 (8 Elite Gen 5), plus an integer-only v73 build for chips whose NPU has no FP16 (8s Gen 4, 8s Gen 3, 7+ Gen 3, 7 Gen 4). A 25-frame 768x512 clip takes about 25 seconds on a Snapdragon 8 Elite; v69 and v73 chips generate at 512x320. Models: timmyboy26/LTX-Video-2B-0.9.8-distilled-QNN.
• Other chips: GGUF models (Q4_0 / Q8_0) run through stable-diffusion.cpp with Vulkan on the GPU, or on the CPU. A 25-frame 512x320 clip takes about 3 minutes on an Exynos 1580 (Galaxy A56) and about 20 minutes on a Tensor G3 (Pixel 8 Pro).
• Open Video Generator, download a model from the Video Generation tab, and choose the model and backend in the settings sheet. Detailed prompts that describe the subject, action, camera and lighting work best.
• The scripts that build the NPU models are in android/tools/ltxv_qnn.

🎵 iOS Live Music Generation

The iOS Music Generator supports Live generation with Magenta Realtime 2: hear audio as it is generated while the app writes it to a local file. Enable Unlimited duration to keep generating until you press Stop, or use a fixed duration. Fixed-duration output uses WAV; unlimited sessions use CAF for long recordings.

🔐 Privacy First

• 100% on-device processing - no internet required for inference
• Zero data collection - conversations never leave your device
• No accounts, no tracking - completely private
• Open-source - fully transparent

⚡ Advanced Capabilities

• GPU/NPU acceleration for fast performance
• Text-to-Speech with auto-readout
• RAG with global memory for enhanced responses
• Import custom models (.task, .litertlm, qnn,.mnn, .gguf)
• Direct downloads from HuggingFace
• 17 language interfaces

🚧 Upcoming Features & Roadmap

Desktop Expansion (Windows & macOS)

• Advanced Vibe Coder — Enhanced code generation with desktop IDE integration, cursor control, and Claude Code–like experiences
• OpenClaw-like Agent — Autonomous tool-using agent framework for complex multi-step tasks
• Advanced Media Generation — Professional-grade image/video synthesis with ComfyUI/Draw Things feature parity and extended controls

Quick Start

1. Download from Google Play or the App Store, get LLM Hub Desktop for your computer, or build from source
2. Open Settings → Download Models → Download or Import a model
3. Select a model and start chatting or generating images

Technology

• Android: Kotlin + Jetpack Compose (Material 3), GenieX SDK (for LLM and VLM inference), WhisperKit (for ASR)
• iOS: Swift + SwiftUI, Run Anywhere SDK, Draw Things (MediaGenerationKit), Apple Foundation Model, whisper.cpp for on-device ASR
• LLM & ASR Runtime: MediaPipe, LiteRT, GenieX SDK (GGUF on Android), WhisperKit (ASR with TFLite + QNN NPU on Android), Llama.cpp (via Run Anywhere SDK on iOS), whisper.cpp on iOS for transcription
• Image & Video Gen: Draw Things (MediaGenerationKit) (iOS), Qualcomm QNN (Android), stable-diffusion.cpp with LTX-Video on the Hexagon NPU (QNN), Vulkan or CPU (Android video)
• Music Gen: Magenta Realtime 2 with MLX/LiteRT (iOS), SoundGen with LiteRT (Android)
• Media Search: EmbeddingGemma 2 for local media indexing and video moment retrieval (Android and iOS)
• Android TTS: Kokoro, Supertonic 3 with ONNX Runtime, and PocketTTS.cpp with an NDK/JNI integration for local voice cloning
• Android Coding Agent: Codex app server in Termux, connected to the app's local model through a loopback inference server

Acknowledgments

• GenieX SDK — GGUF model inference support (credit shown in-app About) ⚡
• WhisperKit — On-device ASR with TFLite + NPU acceleration
• whisper.cpp — On-device ASR/transcription support on iOS
• Run Anywhere SDK — iOS model runtime and LLM execution framework 🚀
• Draw Things — iOS image and video generation engine (MediaGenerationKit) 🎨
• PocketTTS.cpp — Android local speech and voice cloning integration
• Supertone Supertonic 3 — Multilingual on-device TTS and voice presets
• Lightricks LTX-Video — Android video generation model (LTX-Video Open Weights License)
• stable-diffusion.cpp — Android video generation runtime, extended with LTX-Video 0.9 and a QNN backend
• calcuis/ltxv-gguf — GGUF builds of LTX-Video, T5-XXL and the VAE
• Google, OpenAI, Meta, Microsoft, IBM, LiquidAI, Mistral, Primsm ML, HuggingFace — model and tooling contributions

Development Setup

Android local development (Android Studio + Gradle)

git clone https://github.com/timmyy123/LLM-Hub.git
cd LLM-Hub/android
./gradlew assembleDebug
./gradlew installDebug

Android-only local configuration

Setting up Hugging Face Token for Development

To use private or gated models, add your HuggingFace tokens to android/local.properties (do NOT commit this file):

HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
HF_TOKEN_FALLBACK=hf_yyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy

Save and sync Gradle in Android Studio; the app reads BuildConfig.HF_TOKEN and BuildConfig.HF_TOKEN_FALLBACK at build time. HF_TOKEN is used first; the fallback takes over once Hugging Face rejects it (expired or revoked). On iOS, set the same two keys in ios/LLMHub/Secrets.*.xcconfig (copy the .template files).

Dev Premium Flag

To skip ads and unlock all premium features locally without a real IAP purchase, add this to android/local.properties:

DEBUG_PREMIUM=true

Set it back to false before making a production build.

Model License Acceptance

Some models on HuggingFace (especially from Google and Meta) require explicit license acceptance before downloading. When building the app locally:

1. Ensure you have a valid HuggingFace read token in local.properties (see above)
2. For each model you want to download:
   • Visit the model's HuggingFace page (e.g., https://huggingface.co/google/gemma-3n-E2B-it-litert-lm)
   • Click the "Access repository" or license acceptance button
   • Grant consent to the model's license terms
   • Try downloading the model in the app again

Note: This is only required for local development builds. The Play Store version uses different authentication and does not require manual license acceptance for each model.

iOS local development (macOS + Xcode)

Prerequisites

• macOS with Xcode installed (use a version that matches your iOS device version)
• An Apple ID signed into Xcode (free Personal Team works for local device testing)
• iPhone with Developer Mode enabled if you run on real hardware

Build and run on iPhone

1. Clone the repo and open the iOS project:

git clone https://github.com/timmyy123/LLM-Hub.git
cd LLM-Hub
open ios/LLMHub/LLMHub.xcodeproj

2. In Xcode, select target LLMHub → Signing & Capabilities:
   • Set your Team
   • Set a unique Bundle Identifier (for example: com.yourname.llmhub)
   • Keep Automatically manage signing enabled
3. Select your iPhone as the run destination and press Run.

If you use Xcode beta

If your phone is on a newer iOS build and requires Xcode beta support, switch CLI tools:

sudo xcode-select -s /Applications/Xcode-beta.app/Contents/Developer
xcodebuild -version

iOS TTS voice troubleshooting

TTS falls back to a robotic default voice after changing Siri voices

iOS can get into a state where the system TTS voice cache doesn't update after a Siri voice change, causing the app to fall back to the old default voice. To fix:

1. Go to Settings → Accessibility → Spoken Content → Voices
2. Tap your language and select any voice (even a different one)
3. Switch back to your preferred Siri voice or whichever voice you want

This forces iOS to refresh the voice selection and the app will use the correct voice immediately.

Useful iOS dev troubleshooting

• If signing fails, re-check Team + Bundle Identifier in target settings.
• If build cache acts stale, clean DerivedData:

rm -rf ~/Library/Developer/Xcode/DerivedData/LLMHub-*

• Build logs: Report Navigator (Cmd+9)
• Runtime logs: Debug Console (Cmd+Shift+Y)

Contributing

• Fork → branch → PR. See CONTRIBUTING.md (or open an issue/discussion if unsure).

License

• Source code is licensed under the PolyForm Noncommercial License 1.0.0.
• You are free to use, study, and build on this project for non-commercial purposes.
• Commercial use — including distributing the app, charging for it, or monetizing it with ads or IAP — is not permitted without explicit written permission from the author.
• Contact [email protected] for commercial licensing enquiries.

Support

• Email: [email protected]
• Issues & Discussions: GitHub

Notes

• This README is intentionally concise — consult ModelData.kt for exact model variants, sizes, and format details.