Moebius is a tiny AI model that does inpainting—erasing part of an image and filling the gap so seamlessly nothing seems missing—and it runs entirely in a web browser on your own device, with no data sent to any server. Simon Willison ported it to run locally using your machine's graphics chip, meaning no upload, no account, no cloud bill, and your images never leave your computer (project page).
Key facts
- What: Moebius is a small inpainting model claiming far-larger-model quality, and a developer ported it to run on your own machine in a browser tab -- no server, no upload.
- When: 2026-06-22
- Primary source: read the source
Moebius is small enough by modern standards that Willison was able to wrap it into something usable in a browser tab, with the help of a coding assistant—a small story-within-the-story about how quickly capable people can now turn research into working tools (his write-up).
A tiny model running locally matters for three reasons people increasingly care about: privacy, cost, and access. Privacy, because your photos stay on your device instead of being sent to a company's servers. Cost, because there's nothing to pay—no per-image fee, no subscription, just your own hardware doing the work. And access, because a model small enough to run in a browser can reach anyone with a laptop, including people with no fast internet or no budget for cloud services. When AI shrinks to fit on the edge, it stops being a metered utility and starts being more like a feature your device just has.
For years the trend was bigger is better—giant models in giant data centers. The quiet counter-trend is squeezing surprising capability into something small enough to live on your own machine, the way a once-room-sized computer eventually fit in your pocket. Moebius is a small, charming data point on that curve—proof that for some specific, well-defined jobs, you don't need the giant model at all. It hints at a future where many everyday AI features—removing an object, cleaning up a photo, translating a snippet—simply run on your device for free, the way spell-check does today, instead of being metered services you reach across the internet.
The model works by having learned, from huge numbers of images, what tends to go where—that a wall usually continues as a wall, that a face has two roughly symmetric sides, that shadows fall a certain way. When you erase a region, it imagines the most plausible thing that belongs there and paints it in so the edges blend, starting from a patch of random noise and refining it until it agrees with everything around it. That's the same family of technique behind AI image generators, aimed at a smaller, more constrained problem—and doing it well inside a model tiny enough to live in a browser tab is the genuinely hard part.
The honest caveat matters and is easy to overstate past. Moebius's headline claim is that it performs at the level of models many times its size, but that "far-larger-model quality" framing comes from the model's own creators and hasn't been independently verified against named bigger competitors. Tiny models that match big ones on a curated set of examples sometimes fall apart on the messy, varied images of real life, where the big models' extra capacity earns its keep. The right read is: a genuinely impressive, genuinely tiny tool that you can run privately for free today, with a marketing claim about its quality that deserves a healthy pause until outside testing confirms it. Even discounting the boast, "capable image-editing AI that runs free and private in a browser tab" is a real and pleasant thing to have arrived.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)