---
title: Open Model Playground
emoji: 🤗
sdk: gradio
app_file: app.py
pinned: false
---

# Open Model Playground

Colab A100에서는 Qwen3 + SDXL을 실제 실행하고, Hugging Face Space ZeroGPU에서는 Qwen3 + FLUX.2 4-bit 프로파일로 실행하는 강의용 데모입니다.

## Local Colab Profile

```text
TEXT_MODEL_ID=Qwen/Qwen3-4B-Instruct-2507
IMAGE_BACKEND=sdxl
IMAGE_MODEL_ID=stabilityai/stable-diffusion-xl-base-1.0
```

## Space ZeroGPU Profile

```text
TEXT_MODEL_ID=Qwen/Qwen3-4B-Instruct-2507
IMAGE_BACKEND=flux2_4bit
IMAGE_MODEL_ID=diffusers/FLUX.2-dev-bnb-4bit
USE_ZERO_GPU_DECORATOR=1
DEFAULT_USE_REAL_TEXT_MODEL=1
DEFAULT_USE_REAL_IMAGE_MODEL=1
```

## API Endpoints

- `/generate_text`
- `/generate_image`

## Important Notes

- `black-forest-labs/FLUX.2-dev` is gated and uses a non-commercial license.
- The Space profile uses `diffusers/FLUX.2-dev-bnb-4bit`, which contains NF4-quantized components derived from FLUX.2-dev.
- For ZeroGPU, select ZeroGPU hardware in Space settings.
- If real mode fails during recording, turn off `Use real image model` and verify the API path with fallback mode first.