Roost

RinWave

About this app

Roost is a local AI app for Android that lets you discover, download, run, benchmark, and serve AI models directly from your device.

Chat with on-device models, test performance on your own hardware, and start a local network API server for compatible clients. Roost is built for users who want more control over local AI, private experimentation, and mobile model workflows without depending on a cloud chat service for every prompt.

Key features:

• Run local AI models on your Android device
• Chat privately with downloaded on-device models
• Browse and download supported model files
• Manage your local model library and storage
• Benchmark model speed, memory use, and performance
• Start a LAN server with OpenAI-like and Ollama-like chat endpoints
• Use an optional API key for local server access
• Tune inference settings, stop strings, theme, and chat defaults
• Keep prompts and chat activity on your device when using local models

Roost is useful for local AI testing, offline-friendly experimentation, mobile LLM workflows, GGUF model usage, device benchmarking, and running a small private model server on your local network.

The app is designed for early adopters, developers, AI enthusiasts, and privacy-conscious users who want to explore on-device AI from an Android phone. Performance depends on your device, model size, quantization, memory, and selected engine.

Roost does not include model weights in the app package. Model availability, licenses, and download access may depend on third-party model providers such as Hugging Face. Some models may require an access token or may not run well on every device.

Roost is in early development and will continue improving model compatibility, performance, server behavior, and local AI workflows through updates.

What’s new

Replies no longer die when you leave the app. Generation keeps running in the background, and a single notification tells you when the answer is ready — with a Stop button if you change your mind.

• KV cache quantization: up to half the memory at the same context length, or a longer context in the same RAM
• A short first-run tour
• Copy and regenerate on any reply
• Live speed and RAM readouts in chat
• Downloads can be cancelled and clean up after themselves