🔶 Open source on-device AI apps (and counting), no cloud, works offline

⬆️ 5 pts | 💬 4 comments | by rishabh121108

中文全文翻译正在补齐，可先查看下方英文原文。

English · German · Spanish · French · Japanese · Korean · Portuguese · Chinese

On-device is a business decision, not just a technical one.

Every app here runs the model on the phone itself. Nothing goes to a server. That one fact rewrites the economics:

• 🛡️ No compliance wall. No user data in the cloud means no GDPR, HIPAA, or data-residency blocker between you and launch. Put AI into health, finance, and enterprise products, and actually charge for it.
• 💸 $0 marginal cost. No per-token bill, no inference servers. Your margins hold as you scale from 1 to 10 million users.
• 🔒 Private by design. Nothing leaves the device, so there is no cloud dataset to breach, leak, or get audited over.
• ⚡ Instant and offline. Runs on the phone's NPU with no network round-trip, on a plane, a subway, or a factory floor with no signal.

And these are not snippets. Every folder is a finished app you clone and run on a real device today.

Pick any app, clone, and run it on a real device. No ML setup, no model conversion, no C++.

git clone https://github.com/zetic-ai/awesome-on-device-ai-apps.git
cd awesome-on-device-ai-apps

# A free key lets the app pull its NPU-optimized weights on first launch
# (30 seconds, no card): mlange.zetic.ai -> Settings -> Personal Access Token
./scripts/adapt_mlange_key.sh

# Open an app on a REAL device (the NPU isn't in the simulator):
#   Android:  apps/<AppName>/Android    in Android Studio
#   iOS:      apps/<AppName>/iOS        in Xcode
#   Flutter:  cd apps/<AppName>/Flutter && flutter run

Auto-generated from each app's meta.json. Run python3 scripts/generate_catalog.py after adding one.

Jump to: 💬 Language & Text · 👁️ Vision · ❤️ Health & Wellbeing · 🔊 Audio · 📈 Forecasting

🧩 Build your own, by vibe-coding it

Claude Code, Codex, and Cursor will vibe-code you a web app in minutes. Ask them for an app that runs a model on the phone's NPU and they stall, because on-device deployment is not something they know how to do.

That is the gap Melange fills, and it is the easiest way in the world to put AI on-device today. Every app in this repo was built the same way: generate the integration code with Melange, paste it in, done. Copy a use case from here and the on-device feature drops straight into your app, the same vibe-coding loop you already use.

Dropping it into an existing project is about 3 lines:

Android , in build.gradle.kts:

dependencies { implementation("com.zeticai.mlange:mlange:+") }


val model = ZeticMLangeModel(context = this, tokenKey = "YOUR_KEY", modelName = "Team_ZETIC/YOLO26")
val outputs = model.run(inputs)   // NPU-accelerated, on-device

iOS , via Swift Package Manager → https://github.com/zetic-ai/ZeticMLangeiOS.git:

let model = try ZeticMLangeModel(tokenKey: "YOUR_KEY", name: "Team_ZETIC/YOLO26", version: 1)
let outputs = try model.run(inputs: inputs)

Bring your own model: upload it to Melange, it converts and NPU-optimizes automatically, then hands you back a phone-ready build in about an hour, not months of hardware tuning.

This gallery grows by contribution, and the bar is one question: would a stranger clone this and actually use it?

1. Drop your app in apps/<YourApp>/ with Android/ and/or iOS/
2. Add a meta.json (see any existing app) and a README.md
3. Run python3 scripts/generate_catalog.py to add it to the catalog
4. Prove it runs on a real device (demo GIF in the PR)

Full guide → CONTRIBUTING.md. Questions → Discord.

Built by ZETIC · Powered by Melange

If a phone-native AI app made you go "wait, that runs offline?" , then ⭐ star it. It's how the next dev finds it.

App source is Apache 2.0 : use it commercially or privately, however you like. The Melange SDK itself is a proprietary library under the ZETIC Terms of Service.