README.md

---
title: Gemma 4 E2B QAT Mobile Demo
emoji: "\U0001F48E"
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.22.0
app_file: app.py
short_description: Multimodal chat demo of Gemma 4 E2B QAT Mobile
python_version: "3.12"
startup_duration_timeout: 30m
---

# Gemma 4 E2B (QAT Mobile) — Multimodal Demo

A Gradio + ZeroGPU demo running **[`unsloth/gemma-4-E2B-it-qat-mobile`](https://huggingface.co/unsloth/gemma-4-E2B-it-qat-mobile)**, the mobile-optimized (wNa8o8) quantization-aware-trained checkpoint of Google's Gemma 4 E2B model.

## What it does

- Multimodal chat: **text + image** input, streaming text output.
- Optional **thinking mode** (the model reasons step-by-step before answering).
- Adjustable sampling (temperature / top-p / top-k / max tokens).

The model is a 2.3B-effective-parameter multimodal LLM (text, image, audio) quantized to ~2.5 GB via Unsloth's mobile-optimized wNa8o8 scheme, making it runnable on a single ZeroGPU.

## Stack

- **SDK**: Gradio 6 on Hugging Face Spaces
- **Hardware**: ZeroGPU (`zero-a10g`) — free-tier friendly
- **Model loading**: `transformers` `AutoModelForMultimodalLM` + `AutoProcessor`

> The first message after a cold start may take a few extra seconds while ZeroGPU streams the packed weights into VRAM.

## Notes

- Image inputs are placed **before** the text, per the Gemma 4 best practices.
- Default sampling follows the official recommendation: `temperature=1.0`, `top_p=0.95`, `top_k=64`.