Moebius: 0.2B image inpainting model with 10B-level performance

Researchers have developed Moebius, a lightweight image inpainting model with just 0.22B parameters that matches or exceeds the performance of much larger 10B-parameter models like FLUX.1-Fill-Dev while using less than 2% of the parameters and achieving over 15× faster inference speeds. The model achieves this efficiency through a novel Local-λ Mix Interaction (LλMI) block that condenses spatial contexts into fixed-size matrices and an adaptive multi-granularity distillation strategy that transfers knowledge from larger teacher models. Extensive testing on natural and portrait image benchmarks demonstrates that Moebius delivers comparable generation quality to industrial 10B-level models while being practical for deployment on consumer-grade and edge devices.