LLM Hub: free app runs chat, video, music and voice cloning offline on your phone

What's going on

Mexican tech site OneDigital has taken a look at LLM Hub, a free Android and iOS app that runs a long list of AI features entirely on your smartphone. Timmy Qian develops it, and the source code is on GitHub in the timmyy123/LLM-Hub repository. According to the report, you can get the app from Google Play and the App Store. The latest addition is LLM Hub Desktop 0.1.3 for Windows, macOS and Linux.

The app makes three promises: no data collection, no user accounts, no tracking. Every computation happens on the device itself, and nothing goes to the cloud.

More than a chatbot

Most local AI apps only let you chat with a language model you've downloaded. LLM Hub puts more than 15 tools into a single app:

• Chat with web search and RAG memory, so the model can look things up in your own documents before it answers
• Autonomous agent with function calling
• Image generator based on Stable Diffusion 1.5
• Video and music generator
• Transcription with Whisper (Android) or whisper.cpp (iOS)
• Offline translator for more than 50 languages
• Scam detection and photo/video search by description

The interface comes in 17 languages. You can import models in .task, .litertlm, qnn, .mnn and .gguf formats, and downloads come straight from Hugging Face.

Video on your phone: anywhere from 25 seconds to 20 minutes

Video generation is the standout feature. On Android, LLM Hub uses Lightricks' LTX-Video 2B 0.9.8 distilled. This slimmed-down version needs only 8 denoising steps. Models like this start from random noise and refine it step by step into a finished frame, so fewer steps means faster output. You can start from a text prompt or an image.

On Snapdragon chips with a Hexagon NPU, a dedicated processor for AI workloads, the model's core components run as specially compiled binaries. There's a separate build for each chip generation: v69, v73, v75, v79 and v81. NPUs without 16-bit floating-point support get an integer variant. Devices without a compatible NPU run GGUF models quantized to Q4_0 or Q8_0, which are shrunk-down versions that store their numbers at lower precision. These run through stable-diffusion.cpp on the GPU (via Vulkan) or directly on the CPU.

The speed differences are huge:

By OneDigital's math, performance between the best and worst documented cases differs by a factor of almost 50. On iOS, the app uses Stable Video Diffusion for video instead.

An agent that operates your phone

On Android, the autonomous agent can take real actions on your device. It can show interactive OpenStreetMap maps, manage your calendar, set alarms, toggle the flashlight, draft texts and emails, compute cryptographic checksums and do math. It also connects to MCP servers. MCP is an open standard that lets AI assistants plug into external tools such as file access, search or memory. According to the report, you have to approve every single call.

The most unusual part is the Termux integration. Termux is a Linux command line for Android. The agent writes shell commands, shows them in an editor so you can review them, and streams the output live to a terminal. If a command fails, a self-correction loop sends the error message back to the model, which then suggests a fix. In "Vibe Coder" mode, the local model connects to a Codex server running in Termux, and you don't need an OpenAI login. You'll need an ARM64 device running Android 10 or later.

Voice, music and media search

For speech output, Android offers three engines. Kokoro handles local text-to-speech, Supertonic 3 is multilingual and can import voice profiles, and Pocket TTS clones a voice fully locally from 3 to 30 seconds of reference audio. On iOS, Magenta Realtime 2 generates music you can hear while it's still being created. An endless mode keeps recording until you hit stop.

Media search runs locally too. EmbeddingGemma 2 indexes the photos and videos on your device, so you can find them by description or by a sample image. A scene search for videos combines frames and audio from two-second segments, so you can jump straight to the moment you're looking for.

The desktop version

LLM Hub Desktop 0.1.3 is a native app for Windows (x64 and Arm), Apple Silicon Macs (signed and notarized) and Linux (as a .deb package and an AppImage). Besides chatting with local models, you can run Claude Code and OpenClaw on local models, turn prompts into slides and web pages, and expose OpenAI- and Anthropic-compatible APIs. Other programs on your computer can then use your local AI as if it were a cloud service. OneDigital notes that the Windows installer isn't signed yet, so Windows shows a security warning. The project does already sign its builds through SignPath, though.

What you need to try it

For the mobile app, a current Android or iOS device will do. The Termux and Codex features need ARM64 and at least Android 10. If you want videos in a reasonable time, get a device with a recent Snapdragon chip and a Hexagon NPU. On other chips, expect anywhere from a few minutes to 15 minutes or more per clip. The report doesn't list RAM or storage requirements. Since you'll be downloading models, plan on several gigabytes of free space. Some Google and Meta models require you to accept their license on Hugging Face first. The desktop version runs on Windows, Apple Silicon Macs and Linux.

Our take

OneDigital calls LLM Hub probably the most complete free local AI app in the official app stores. Given everything it does, that's believable, but no comparison tests back it up. What stands out most is that video generation and voice cloning now work on an ordinary smartphone with no internet connection.

Two things take the edge off. The first is the license. PolyForm Noncommercial 1.0.0 lets you use, study and modify the code. To distribute or sell it, or to make money from it through ads or in-app purchases, you need the author's written permission. Strictly speaking, that makes LLM Hub source-available with restrictions, not open source in the classic sense. The second is performance, which varies wildly and takes a lot of patience on many devices. For serious video work, the app is no substitute for a computer with a powerful GPU. The roadmap promises an expanded Vibe Coder with IDE integration and an agent for multi-step tasks, but there's no date for either. Still, as a private AI lab that fits in your pocket, LLM Hub is an exciting project.