Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code
Simon Willison ported the Moebius 0.2B image inpainting model to run in a browser using WebGPU, with assistance from Claude Code. The model, originally requiring PyTorch and NVIDIA CUDA, was converted to ONNX format and deployed on Hugging Face. The demo is available at simonw.github.io/moebius-web/. The project was a side effort while working on a major Datasette feature. Willison first used Claude.ai to research feasibility, then set Claude Code to work in a /tmp folder with prepared materials. Claude Code maintained notes.md and plan.md files, and after several iterations, produced a working browser-based demo. The 1.24GB of converted ONNX weights were published to huggingface.co/simonw/Moebius-ONNX. The frontend code is hosted on GitHub Pages. Willison noted that each page load downloaded ~1.3GB of model weights, and he explored service workers for caching.
Demonstrates that small image models can be ported to run entirely in the browser via WebGPU.