Files read on Hugging Face: 2026-10-31

MOSS-Image on Mac: files, RAM and published times

MOSS-Image is an Apache 2.0 text-to-image model from OpenMOSS with a 5B diffusion transformer, a Qwen3.5 9B text encoder and the FLUX.2 VAE. This page covers the SFT checkpoint, whose card defaults to 50 steps.

OpenMOSS · Updated 2026-10-31
Answer

MOSS-Image runs on a Mac with 32 GB or more. A published run on M1 Max · 32 GB used the Original SFT safetensors build (source, 2026-10-18).

Smallest Mac with a published run
32 GBM1 Max · 32 GB · Original SFT safetensors
Smallest complete download
13.20 GBUD-Q4_K_XL GGUF
Files read
2026-10-31Hugging Face

Apache 2.0. Commercial use: allowed. Excluded territories: none listed.

Table 1

What do you download?

Only builds with every required component and an exact probed file size are shown.

Original SFT safetensors

safetensors · BF16 (original weights) · 27.20 GB complete · Recommended

Mac routes: MOSS-Image MPS path

Exact component files for Original SFT safetensors
FileComponentRequiredDownloadSource
transformer/diffusion_pytorch_model-00001-of-00003.safetensorsOpenMOSS-Team/MOSS-Image-SFTMOSS-Image transformerditYes4.99 GBHugging FaceRead 2026-10-31
transformer/diffusion_pytorch_model-00002-of-00003.safetensorsOpenMOSS-Team/MOSS-Image-SFTMOSS-Image transformerditYes4.98 GBHugging FaceRead 2026-10-31
transformer/diffusion_pytorch_model-00003-of-00003.safetensorsOpenMOSS-Team/MOSS-Image-SFTMOSS-Image transformerditYes0.19 GBHugging FaceRead 2026-10-31
text_encoder/model-00001-of-00004.safetensorsOpenMOSS-Team/MOSS-Image-SFTQwen3.5 9B text encodertext_encoderYes4.97 GBHugging FaceRead 2026-10-31
text_encoder/model-00002-of-00004.safetensorsOpenMOSS-Team/MOSS-Image-SFTQwen3.5 9B text encodertext_encoderYes4.99 GBHugging FaceRead 2026-10-31
text_encoder/model-00003-of-00004.safetensorsOpenMOSS-Team/MOSS-Image-SFTQwen3.5 9B text encodertext_encoderYes4.95 GBHugging FaceRead 2026-10-31
text_encoder/model-00004-of-00004.safetensorsOpenMOSS-Team/MOSS-Image-SFTQwen3.5 9B text encodertext_encoderYes1.87 GBHugging FaceRead 2026-10-31
vae/diffusion_pytorch_model.safetensorsOpenMOSS-Team/MOSS-Image-SFTFLUX.2 VAEvaeYes0.25 GBHugging FaceRead 2026-10-31

Build sources: MOSS-Image Apple silicon guide.

UD-Q4_K_XL GGUF

gguf · UD-Q4_K_XL · 13.20 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for UD-Q4_K_XL GGUF
FileComponentRequiredDownloadSource
MOSS-Image-SFT-UD-Q4_K_XL.ggufunsloth/MOSS-Image-SFT-GGUFMOSS-Image transformerditYes3.42 GBHugging FaceRead 2026-10-31
Qwen3.5-9B-Q8_0.ggufunsloth/Qwen3.5-9B-GGUFQwen3.5 9B text encodertext_encoderYes9.53 GBHugging FaceRead 2026-10-31
vae/diffusion_pytorch_model.safetensorsunsloth/MOSS-Image-SFT-FP8FLUX.2 VAEvaeYes0.25 GBHugging FaceRead 2026-10-31

Build sources: MOSS-Image-SFT GGUF card · stable-diffusion.cpp MOSS-Image docs.

Q8_0 GGUF

gguf · Q8_0 · 15.26 GB complete

Mac routes: stable-diffusion.cpp

Exact component files for Q8_0 GGUF
FileComponentRequiredDownloadSource
MOSS-Image-SFT-Q8_0.ggufunsloth/MOSS-Image-SFT-GGUFMOSS-Image transformerditYes5.48 GBHugging FaceRead 2026-10-31
Qwen3.5-9B-Q8_0.ggufunsloth/Qwen3.5-9B-GGUFQwen3.5 9B text encodertext_encoderYes9.53 GBHugging FaceRead 2026-10-31
vae/diffusion_pytorch_model.safetensorsunsloth/MOSS-Image-SFT-FP8FLUX.2 VAEvaeYes0.25 GBHugging FaceRead 2026-10-31

Build sources: MOSS-Image-SFT GGUF card · stable-diffusion.cpp MOSS-Image docs.

Table 2

Does it fit your Mac's RAM?

Published names the exact measured machine. est. is a memory calculation. unknown means the run verdict is not established, not that the model fails.

Run verdicts by Mac memory tier
Mac RAMComplete buildDownloadRun verdictBasis
8 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBDoes not fitest.
16 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBunknownunknown
24 GBOriginal SFT safetensorsBF16 (original weights) · MOSS-Image MPS path27.20 GBunknownunknown
32 GBOriginal SFT safetensorsBF16 (original weights) · MOSS-Image MPS path27.20 GBRuns · PublishedPublishedApple M1 Max · 32 GB RAM
48 GBOriginal SFT safetensorsBF16 (original weights) · MOSS-Image MPS path27.20 GBRuns · Published on M1 Max · 32 GBPublishedApple M1 Max · 32 GB RAM
64 GBOriginal SFT safetensorsBF16 (original weights) · MOSS-Image MPS path27.20 GBRuns · Published on M1 Max · 32 GBPublishedApple M1 Max · 32 GB RAM
96 GBOriginal SFT safetensorsBF16 (original weights) · MOSS-Image MPS path27.20 GBRuns · Published on M1 Max · 32 GBPublishedApple M1 Max · 32 GB RAM
128 GBOriginal SFT safetensorsBF16 (original weights) · MOSS-Image MPS path27.20 GBRuns · Published on M1 Max · 32 GBPublishedApple M1 Max · 32 GB RAM

Every image model, ranked by memory: 8 GB Mac · 16 GB Mac · 24 GB Mac · 32 GB Mac · 64 GB Mac

Table 3

Does it fit your NVIDIA GPU's VRAM?

Runs · Stated means the linked model card, app maker or NVIDIA says a GPU with that much VRAM runs the model; Runs · Published names the card of a dated run. Fits with offload means the main model weights fit in VRAM with a 15% margin but the full set does not: the text encoder and VAE run from system RAM, which is slower. Exceeds VRAM means even the main weights are too large for the card. Mac runs never count here.

Run verdicts by NVIDIA VRAM tier
NVIDIA VRAMComplete buildDownloadRun verdictBasis
8 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBFits with offloadest.Main weights 3.42 GB: fit with a 15% margin
12 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBFits with offloadest.Main weights 3.42 GB: fit with a 15% margin
16 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBunknownunknown
24 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBunknownunknown
32 GBUD-Q4_K_XL GGUFUD-Q4_K_XL · stable-diffusion.cpp13.20 GBunknownunknown

Every image model, ranked by memory: 8 GB of VRAM · 12 GB of VRAM · 16 GB of VRAM · 24 GB of VRAM

How does it rank in blind votes?

No independent score yet.

Table 4

What published Mac times are available?

Third-party results keep their original machine and settings. ModelFit does not run these benchmarks.

Published Mac generation measurements
Machine and RAMSettingsAppResultSource
Apple M1 Max · 32 GB RAM1024x1024 · 20 steps · BF16 · Official repository record for the SFT checkpoint through the experimental MPS path, seed 42, torch 2.14.1 and diffusers 0.41.0. A full recorded generation call, not a controlled throughput or peak-memory measurement. The date is the commit date of the Mac guide; the run date is not stated.MOSS-Image MPS pathSuccess320.4 sPeak memory unknownMOSS-Image Apple silicon guidePublished 2026-10-18 · Grade A

Which Mac software routes are listed?

  • MOSS-Image MPS path

    Official · experimental

    Formats: safetensors. Memory policy: components can be unloaded between stages.

    Experimental PyTorch MPS path in the official MOSS-Image repository. It stages the Qwen text encoder and the DiT so only one is resident at a time, keeps the text encoder in BF16 and the VAE in FP32, and applies no quantization.

    Set up and launch from the cloned MOSS-Image repository

    MOSS_PYTHON=python3.12 bash scripts/setup-mac.sh .venv/bin/hf download OpenMOSS-Team/MOSS-Image-SFT --local-dir models/MOSS-Image-SFT bash scripts/gradio-mac.sh

    Sources: MOSS-Image Apple silicon guide

  • stable-diffusion.cpp

    Official

    Formats: gguf, safetensors. Memory policy: component stages stay resident.

    C/C++ runner with Metal on Mac and CUDA on NVIDIA. It reads GGUF and safetensors, and its docs keep a page per model family. Resident size is the sum of the weights, an upper bound.

    Sources: stable-diffusion.cpp README, backends

What does the official configuration say?

"Recorded real-weight validation is on an M1 Max with 32 GB of unified memory, with SFT. Other chips, lower-memory systems, Base and older dependency versions need testing."

Source: MOSS-Image Apple silicon guide · Read 2026-10-31.

What known issues should you check?

  • Known issue 1

    The Mac path is experimental. CUDA offload modes and Compile DiT do not apply to it, and the authors did not validate the full ComfyUI path on a Mac.

    Source: MOSS-Image Apple silicon guide · Read 2026-10-31

  • Known issue 2

    The recorded Mac times are end-to-end records, not controlled throughput or peak-memory measurements, so no Mac peak memory is listed.

    Source: MOSS-Image Apple silicon guide · Read 2026-10-31

  • Known issue 3

    The GGUF files hold only the transformer. The VAE and tokenizer come from unsloth/MOSS-Image-SFT-FP8 and the text encoder from unsloth/Qwen3.5-9B-GGUF.

    Source: MOSS-Image-SFT GGUF card · Read 2026-10-31

Frequently asked questions

Can I run MOSS-Image on a Mac?

Yes. The official MOSS-Image repository records SFT runs on an M1 Max with 32 GB through its experimental PyTorch MPS path. A 1024x1024 image in 20 steps at BF16 took 320.362 s. The authors have not tested smaller Macs.

How does MOSS-Image fit in Mac memory?

The MPS path stages the Qwen text encoder and the DiT so only one of them is resident at a time, and it applies no quantization. The exact file sizes are listed below; the only published Mac run used 32 GB of unified memory.

Does MOSS-Image have GGUF files?

Yes. unsloth publishes transformer-only GGUF files for stable-diffusion.cpp builds from master-975 on. They need the Qwen3.5-9B Q8_0 GGUF text encoder and the VAE from unsloth/MOSS-Image-SFT-FP8. No published Mac run with a GGUF build is listed yet.

Sources