# Fish Anything — development

## Models (each ≤32B)

| Model | Params | Role |
|-------|--------|------|
| [openbmb/MiniCPM4.1-8B](https://huggingface.co/openbmb/MiniCPM4.1-8B) | 8B | Bait appeal, fish lore, reel scoring, narration |
| [black-forest-labs/FLUX.2-klein-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B) | 4B | Bait & fish sprites (fallback) |
| [black-forest-labs/FLUX.2-klein-9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B) | 9B | Bait & fish sprites (primary) |

Text and image run on separate Modal containers (A10G + A100). The Hugging Face Space is a CPU Gradio host that calls Modal remotely.

## Tech

- **Frontend:** `gr.Server` + static HTML/CSS/JS
- **Backend:** FastAPI routes on the Gradio server; per-session game state
- **Modal:** `modal deploy modal_app.py` — `TextGenerator` + `ImageGenerator`
- **Warmup:** Modal containers wake on Space boot and page load

## Space secrets (live play)

| Secret | Value |
|--------|-------|
| `MODAL_TOKEN_ID` | From `modal token new` or [Modal settings](https://modal.com/settings) |
| `MODAL_TOKEN_SECRET` | Paired secret |
| `LLM_MODE` | `modal` |
| `IMAGE_MODE` | `modal` |

Deploy Modal once: `modal deploy modal_app.py` (needs `huggingface-secret` for FLUX weights).

## Local dev

```bash
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
LLM_MODE=mock IMAGE_MODE=mock python app.py
# http://127.0.0.1:7860
```