README.md

---
title: TensorBend
emoji: ⚡
colorFrom: gray
colorTo: green
sdk: static
pinned: false
---

# TensorBend

Run LLMs entirely in your browser. Weights are fetched as raw SafeTensors from HuggingFace and loaded directly into WebGPU compute buffers. No ONNX, no server.

**Requires:** Chrome/Edge with WebGPU support (macOS, Windows, ChromeOS). Apple Silicon recommended for best performance.

**Supported models:** Qwen3.5 family (0.8B, 2B, 4B, 9B) — INT4 quantized via AutoRound.