Moebius: a 0.2B image inpainting model that rivals the 11.9B FLUX.1-Fill-Dev

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

Huazhong University of Science and Technology and VIVO AI Lab released Moebius, a 226M-parameter image inpainting model that matches or beats the 11.9B FLUX.1-Fill-Dev across 6 benchmarks while using less than 2% of the parameters. It runs at 26 ms per step, giving a >15× total speedup over 10B-level models. The key trick: reformulating self- and cross-attention into fixed-size linear matrices to dodge quadratic compute, plus an adaptive multi-granularity distillation that transfers knowledge from the PixelHacker teacher entirely in latent space. I'd hold off popping champagne—the paper compares mainly against FLUX.1-Fill-Dev and SD3.5 Large-Inpainting, and the post doesn't disclose real throughput or memory numbers on consumer GPUs. Code and paper are public.

Why it matters: Moebius matches 10B-level inpainting models with only 0.2B parameters by replacing attention with fixed-size linear matrices, sidestepping quadratic compute growth. H and K are solid, but the paper-only release lacks a product or open-source spark, so R is weak — right at the ...