Zyra

Zyra is a small Gradio app for plant ID and plant-care questions. Upload a photo, type a question, or record your voice. It uses a local GGUF vision-language model, Whisper for speech input, Kokoro for speech output, and Trefle when a botanical match is available.

Demo

demo.mov

The demo is committed as demo.mov in the repo root. To get an inline player on GitHub, drag demo.mov into the README via GitHub's web editor — it will replace this link with an https://github.com/user-attachments/... URL that embeds the video.

Tech

• Python 3.14
• Gradio for the app UI
• llama.cpp for local GGUF inference
• Quantized Gemma 4 (12B / E4B) for plant ID, image understanding, and care answers
• Whisper Large v3 Turbo for speech-to-text
• Kokoro 82M for text-to-speech
• Trefle API for botanical lookup and tool calling

The plant reasoning model is local. Use quantized gemma-4-12B when you want better answers and can wait a bit longer. Switch to quantized gemma-4-E4B when you want faster responses and can accept lower accuracy.

Run

cp env.example .env
uv run python main.py

After it starts, open http://127.0.0.1:7860/.

Set TREFLE_TOKEN in .env if you want verified plant data. Without it, Zyra still answers from the model, but skips the Trefle lookup.

Notebook

Open main.ipynb if you prefer running the app cell by cell.