README.md

---
title: Falcon-Perception-0.6B WebGPU
emoji: 🦅
colorFrom: indigo
colorTo: pink
sdk: static
pinned: false
license: apache-2.0
short_description: Open-vocab detection + segmentation, all in the browser
models:
  - tiiuae/Falcon-Perception
  - onnx-community/falcon-perception-onnx-webgpu
---

# 🦅 Falcon-Perception-0.6B WebGPU

A browser demo for **[tiiuae/Falcon-Perception](https://huggingface.co/tiiuae/Falcon-Perception)** — a 0.6B open-vocabulary VLM that turns natural-language queries into bounding boxes and pixel-accurate segmentation masks, running fully client-side via WebGPU + ONNX Runtime Web.

[![Model](https://img.shields.io/badge/🤗%20Model-tiiuae%2FFalcon--Perception-yellow)](https://huggingface.co/tiiuae/Falcon-Perception)
[![Weights](https://img.shields.io/badge/🤗%20ONNX%20Weights-onnx--community%2Ffalcon--perception--onnx--webgpu-blue)](https://huggingface.co/onnx-community/falcon-perception-onnx-webgpu)

## What's inside

- **Detection** — draw bounding boxes for any natural-language query ("athletes", "the runner in front", "mangoes").
- **Segmentation** — pixel-accurate masks via the AnyUp upsampler, all in-browser.
- **Tracker (preview)** — HUD-style reticles on video. Limited by VLM latency between detections; see the in-space disclaimer.

## How it runs

2.4 GB of ONNX weights are fetched once on first visit, then cached by your browser — no backend, no API keys, no network round-trip after load. Multi-threaded WASM is enabled via `coi-serviceworker`.