README.md

---
title: System One Arcade
emoji: 🟣
colorFrom: purple
colorTo: gray
sdk: docker
app_port: 7860
suggested_hardware: l4x1
models:
  - LiquidAI/d1-3B
short_description: Eleven camera, drawing and text demos of d1-3B
pinned: false
---

# System One Arcade

Eleven demos of `d1-3B`. Each one asks the model a few typed questions about a camera frame,
a drawing or a message, and reads the answers from one forward pass.

| demo | what it shows |
|---|---|
| Hand Flappy | open your palm to fly up, close it to drop |
| Pong vs the Model | the model plays from a text description of the game |
| Rep Counter | counts squats, jumping jacks or arm raises |
| Keep It Clean | pixelates the middle finger, 0.4 s behind the camera |
| Find It | show the camera each object it asks for |
| Quick Draw | it guesses your drawing among 16 words |
| Red Light, Green Light | move on green, freeze on red |
| Cowboy Duel | two players, first finger gun wins |
| Live Triage | six answers about a support message as you type |
| Fit Check | live outfit score with purikura-style doodles |
| Ask Anything | your own yes/no questions about the camera image |

Camera frames go to the model on this Space's GPU and are not stored.

## Run it locally

Here every frame makes a round trip to the Space's server, about 0.2 s, which the games that react
to each frame (Flappy, Pong, Duel, Red Light) can feel. On your own machine a decision comes back
in tens of milliseconds.

You need an Apple silicon Mac, or Linux (WSL2 works) with an NVIDIA GPU of 8 GB or more, and
Python 3.12 or newer.

```bash
git clone https://huggingface.co/spaces/LiquidAI/system-one-arcade
cd system-one-arcade
python3 -m venv .venv && source .venv/bin/activate
```

On an Apple silicon Mac (PyTorch on mps):

```bash
pip install -r requirements-mac.txt
python game/server.py
```

On an NVIDIA GPU (PyTorch, CUDA graphs):

```bash
pip install -r requirements-nvidia.txt
python game/server.py --compile
```

Open the address it prints; it carries a token, so other machines on your network cannot use the
server. The first start downloads the model (6.2 GB) and warms up for a minute or two; if the
download asks for a login, run `hf auth login` first.
`python game/server.py --help` lists the options (port, model, frame size).

**Settings:**
- **Visibility:** private until launch, then public. The server has no token, so anyone who can open
  the Space can use its GPU;
- **Hardware:** an NVIDIA GPU with bfloat16, an L4 or larger;
- **Secret:** `HF_TOKEN`, a token that can read the model.