| > 📱 **Now on Google Play — [Fancy AI](https://play.google.com/store/apps/details?id=com.mrj.fancyai)** | |
| > On-device AI image generation, optimized for Qualcomm NPU phones. | |
| --- | |
| license: other | |
| language: | |
| - en | |
| tags: | |
| - gguf | |
| - llama.cpp | |
| - android | |
| - fancy-ai | |
| - on-device | |
| pretty_name: Fancy AI — Add-on Catalog | |
| --- | |
| # Fancy AI — Add-on Catalog | |
| Add-on catalog for the **Fancy AI** Android app — an on-device AI companion app powered by | |
| `llama.cpp`. The app ships small and streams optional models on demand from this catalog. | |
| - App & issues: https://github.com/Mr-J-369/Fancy-Ai | |
| - This repo hosts **`manifest.json`** — the catalog the app reads. The model weights themselves | |
| are **not re-hosted here**; the catalog links directly to the original creators' repos | |
| (Qwen, bartowski), so downloads come straight from the source. | |
| ## How to use it (in the app) | |
| **Settings → On-Device Model → Browse Add-ons.** The app reads `manifest.json`, lists each | |
| component, and lets you **download + activate** in one tap: | |
| - `agent` → **Use as Root's brain** (the tiny model that handles background housekeeping). | |
| - `chat` → **Use as chat model** (your main conversational model). | |
| ## Current catalog | |
| │ Component │ Type │ Model │ Size │ License │ | |
| │ Root Agent — Qwen2.5 1.5B │ agent │ Qwen2.5-1.5B-Instruct │ 1.12 GB │ Apache-2.0 │ | |
| │ Chat — Llama 3.2 3B │ chat │ bartowski Llama-3.2-3B │ 2.02 GB │ Llama 3.2 │ | |
| │ Chat — Qwen2.5 3B │ chat │ Qwen2.5-3B-Instruct │ 2.10 GB │ Qwen │ | |
| ## Manifest format | |
| `manifest.json` is a single JSON file with a `components` array. Each component has: | |
| `id`, `type` (`agent` | `chat` | …), `name`, `description`, `quant`, `minRamMb`, `url` | |
| (a direct `…/resolve/main/<file>.gguf` link), `sizeBytes`, and `sha256`. Multi-file components | |
| (e.g. a vision model + projector) use a `files` array instead of a single `url`. | |
| ## Licenses | |
| Each model retains its **original license** (Apache-2.0 for Qwen 0.5B; the Llama 3.2 and Qwen | |
| community licenses for the 3B models). Weights are downloaded directly from the creators' repos; | |
| this repo only distributes the catalog metadata. See each linked model card for full terms. | |