Moebius, a 0.2B image inpainting model from Hustvl, now runs entirely in a browser via WebGPU. Simon Willison converted the original PyTorch and CUDA-only model to ONNX format, uploaded 1.24GB of converted weights to Hugging Face at huggingface.co/simonw/Moebius-ONNX, and deployed a working demo at simonw.github.io/moebius-web. The tool lets you open any image, mask regions, and have the model reconstruct what belongs there, all client-side.
Claude Code did the heavy lifting. Willison ran it in a terminal while waiting on a separate Codex Desktop session to finish a Datasette refactor, treating this as a parallel side project. The architecture choice matters: Claude recommended ONNX Runtime Web on the WebGPU backend, one layer below Transformers.js, which is what made browser execution viable. Willison seeded the agent with cloned repos of Moebius, its weights, Transformers.js, and ONNX Runtime before prompting it to write a plan.md, commit early and often, and maintain a notes.md throughout.
The full write-up is worth reading for the workflow, not just the result. Willison's method of using Claude.ai first to assess feasibility, saving that output as research.md for Claude Code to ingest, and iterating by pasting browser error screenshots back into the agent is a concrete, repeatable process. The notes.md and plan.md artifacts the agent produced are linked and show how an agent documents its own reasoning across a multi-hour session.
[READ ORIGINAL →]