Moebius image repair gets a 247MB community GGUF

Image-inpainting experiments have a new compact download: cstr has published Moebius in GGUF form, with the smallest released weight file weighing about 247 MB. The public repository appeared on November 2 at 15:48:37 UTC. Its card says the artifacts are exact copies of previously qualified files, with no requantization performed for publication.

The four files span F32 at about 905 MB, F16 at 485 MB, Q8_0 at 288 MB and Q4_K-mix at 246,888,384 bytes. This is a community conversion of existing Moebius weights for image repair, rather than a newly trained upstream model.

A specific engine, not a universal GGUF download

The card says these files run only in CrispEmbed’s native inpainting engine, behind the default-off CRISPEMBED_MOEBIUS=1 gate. Non-F32 variants also require CRISPEMBED_INPAINT_QUANT=1. Each needs a separate SDXL VAE companion; the smallest Moebius variant pairs with the Q8_0 VAE, listed at roughly 89 MB. The 247 MB headline therefore describes one weight file, not the whole installation or a measured memory requirement.

The reviewed public card points to CrispEmbed documentation but does not supply a runtime download or installation command. Minimum hardware, supported operating systems and measured inference speed are not established there. AITrending has not run the files.

Why the smallest file uses mixed precision

The publisher says plain Q4_K failed its real-photo acceptance threshold. The released mixed variant keeps every attention-block tensor at Q8_0 and reports a minimum cosine similarity of 0.9449 on its photo comparison. Synthetic cases are compared with the original PyTorch model; real-photo masked regions are compared with native F32 output.

These are the publisher’s fidelity checks, not an independent benchmark or proof of better visual quality. The card also preserves conflicting upstream weight-license records, Apache-2.0 in the official README and MIT in the author’s Hugging Face card, without claiming a dual license. The practical news is the availability of experimental precision choices with explicit tradeoffs and limits.