Tencent's Moebius packs inpainting into 0.2B params

Tencent has released Moebius, an image inpainting framework that fits in roughly 0.2 billion parameters yet claims output quality approaching models more than an order of magnitude larger. For a task that usually leans on heavy diffusion backbones, that is an unusual bet on efficiency over raw scale.

Inpainting—filling in or replacing parts of an image so the result looks seamless—is one of the more practical corners of generative imaging, used for object removal, retouching, and background edits. The project page (hustvl.github.io/Moebius) positions the model as a compact alternative to the 10B-class systems that dominate the category.

Why it matters

A model this small changes where the work can happen. Developer Simon Willison ported Moebius to run directly in the browser with the help of Claude Code, demonstrating that the framework is light enough for client-side inference rather than a GPU server.

That has real consequences:

• Privacy: images never leave the user's device.
• Cost: no inference servers to provision or scale.
• Latency: edits run locally without a network round trip.

The release is offered under a non-standard license, so teams will want to check the terms before building on it. But as a proof point, Moebius suggests there is still meaningful headroom for shrinking specialized image models without giving up much in quality.