README.md

metadata

title: Rock, Paper, d1
emoji: ✊
colorFrom: purple
colorTo: yellow
sdk: static
app_file: index.html
pinned: false
short_description: Rock Paper Scissors refereed by d1-3B, on your GPU
models:
  - onnx-community/d1-3B-ONNX
  - LiquidAI/d1-3B

Rock, Paper, d1

Rock, Paper, Scissors against a random bot, refereed by Liquid AI's d1-3B, a 3B decision model that reads your hand from the webcam. It runs entirely in your browser on WebGPU with Transformers.js: your camera never leaves the page.

• Each read is one forward pass over one webcam frame and one typed question ("What is their hand showing?": an open palm, a peace sign, a closed fist, or no hand). The answer is a softmax over the options, as in d1's own system_one. Nothing is generated.
• About 1.3 s per read on an M3 Pro (128 image tokens, 8-bit weights). Between rounds the bars show d1 reading your hand live.
• A round counts when d1's top gesture is at least 45% and leads the runner-up by 10 points; otherwise you replay it.
• Desktop Chrome or Edge with WebGPU. About 3.8 GB of model files on first load, cached after.

Model: LiquidAI/d1-3B (LFM Open License v1.0), converted to ONNX at onnx-community/d1-3B-ONNX.