README.md

---
title: Amigo
emoji: 👋
colorFrom: yellow
colorTo: green
sdk: gradio
sdk_version: 6.16.0
python_version: '3.12'
app_file: app.py
pinned: false
short_description: A private, local voice companion that knows you
models:
- pebeto/amigo-lora
---

# 👋 Amigo

A privacy-first voice companion that runs on small models and knows the person it talks to. Built for the [Build Small Hackathon](https://huggingface.co/build-small-hackathon).

My father uses Gemini. It confuses him, and it always needs the internet. Amigo works differently. He presses one big button and talks. It answers in his language, remembers him, keeps his data on the device, and goes online only for questions that need current facts.

## How to use it

1. Open **Mi perfil** and describe the person (the example shows the format).
2. Press the big green button, speak, then press the red button to stop.
3. Amigo replies out loud and shows the words in large text.

## What runs here

```
🎤 mic → faster-whisper (es) → Qwen3.5-2B + amigo LoRA (llama.cpp)
           ├─ profile + Chroma memory (local)
           └─ DuckDuckGo search (only for current facts)
       → Piper TTS (es) → 🔊 speaker
```

This Space runs the small, fine-tuned **amigo-2b** ([pebeto/amigo-lora](https://huggingface.co/pebeto/amigo-lora) on Qwen3.5-2B), chosen so replies stay quick on a CPU. Models download from the Hub on first boot, so the first load takes a few minutes.

The language is configurable in `i18n.py` (Spanish by default, English included). Set the `MODEL_KEY` variable to `qwen3.5-4b` to run the larger model instead.