MOSS-Image OrbitQuant W4A8

Model Oct 17, 2026

WaveCut's community quantization of MOSS-Image-SFT, packing the image transformer and text encoder into an 8.8 GB bundle for a documented Linux and ComfyUI workflow.

Visit primary link 2 sources attached

WaveCut's MOSS-Image chart comparing generation time, peak memory and checkpoint size for BF16 and OrbitQuant profiles on one RTX 5090. — Author-reported RTX 5090 comparison from the English quantization card. These are WaveCut's measurements, not an AITrending benchmark or results guaranteed on other GPUs. Full chart and measurement footnote preserved. Source: WaveCut / MOSS-Image OrbitQuant

multimodal self-hosted

What happened lately

MOSS-Image OrbitQuant W4A8 timeline

Ecosystem Oct 17, 2026

MOSS-Image gets an 8.8 GB community build with a ComfyUI recipe

Community reports 4 sources

This community build applies OrbitQuant four-bit weights and eight-bit activations to the MOSS-Image-SFT transformer and Qwen3.5 text encoder. The author’s English card documents an 8.8 GB checkpoint bundle, CLI and Python entry points, and ComfyUI-OrbitQuant 0.8.0 profiles.

Performance figures are the author’s measurements on one RTX 5090, not independent tests or guarantees for other GPUs. The card reports additional failures for Arabic and Korean text versus BF16 and explicitly says OpenMOSS did not create, endorse or validate the conversion. It declares Apache-2.0 licensing. Repository creation on October 16 dates this build, not the original upstream model.