AI Playground
The SafeCircle AI Playground for Android and the web playground - what they are, which models and runtimes they support, how to get them, and how they handle your data.
SafeCircle publishes two playgrounds where you can try its risk-detection models yourself. Neither one monitors anything, and neither is part of the parental-control product.
Experimental tools
Both playgrounds are for exploring and testing models. Results can be wrong. Do not use them as the only safeguard for a child.
SafeCircle AI Playground for Android
An Android app for running AI models directly on your phone: chat with them, test them and benchmark them. It is based on Google's AI Edge Gallery and is released under the Apache 2.0 license.
What you can do
| Feature | What it does |
|---|---|
| AI Chat | Multi-turn conversations with an on-device language model |
| Prompt Lab | Single prompts with templates and settings |
| Ask Image / Ask Audio | Image and audio questions with models that support them |
| Safety Detection | Paste a conversation and see how the Horizon models classify it |
| Vigil | Score a whole conversation with the Vigil research model |
| Agent Chat | Let a model use built-in skills and MCP servers |
| Benchmark | Measure latency and throughput on your hardware |
| Model Manager | Browse, download and delete models, or import your own from storage or a URL |
Models
| Model | Notes |
|---|---|
| Horizon Mobile INT4 | Fastest safety model (Gemma 3 1B, 4-bit) |
| Horizon Mobile INT8 | Higher accuracy for the same task (Gemma 3 1B, 8-bit) |
| Horizon Edge | General-purpose agent model based on Gemma 4 E2B, for phones with 8 GB RAM or more |
| Vigil | Compact multilingual encoder, about 1.3 GB, runs on the CPU |
Models run on Google AI Edge runtimes (LiteRT and LiteRT-LM). You can choose CPU, GPU or NPU where the phone and model support it, with an automatic fall back to CPU. Some gated models need a Hugging Face sign-in to download.
Requirements and how to get it
- Android 11 or newer. 4 GB of RAM or more is recommended for larger models.
- Models are large, so use Wi-Fi and keep free storage.
- The source code and releases are on GitHub in the
safecircleia/ai-playground-androidrepository.
Privacy
Once a model is downloaded, the text you test is processed on your phone and is not sent to SafeCircle. The network is used to download models and by features you choose to use, such as gated Hugging Face downloads, Wikipedia lookups or MCP servers. No analytics or crash-reporting SDK is wired into the app build.
Which one should I use
| I want to | Use |
|---|---|
| Try the models on my own phone, including Vigil | Android app |
| Compare behaviour quickly without installing anything | Web playground (beta) |
| Protect and monitor a child's device | The SafeCircle child app, not a playground. See Set up the child app |