All stories
MOSS-Image-SFT MLX Ecosystem Community reports

MOSS-Image goes fully MLX on Macs, with new 5.8GB Q3 and 7.5GB Q4 packs

MFlux merged a complete native MLX pipeline, including the Qwen3.5 text encoder, on October 22. Community Q3/Q4 packs followed, but the currently published PyPI wheel still lacks the new command.

Sources (6)

Why now Pull request 914 merged October 22 at 00:03 UTC, replacing the earlier fork proposal. Complete MFlux community Q3 and Q4 repositories appeared around 00:53 UTC, a new integration and packaging event rather than another launch of the original MOSS model.

Storyline / Report 2 of 2

MOSS-Image's MLX port and native MFlux integration

Latest report
Full timeline 2 reports
  1. Community reports
    MOSS-Image-SFT gets a 4-bit Apple Silicon port
  2. Community reports
    MOSS-Image goes fully MLX on Macs, with new 5.8GB Q3 and 7.5GB Q4 packs (current report)
MFlux's generated reference image shows a vintage red bicycle beside an ivy-covered stone wall in a sunlit alley.
Project-published MOSS-Image-SFT regression reference: 640 by 368, 20 steps, seed 7 and 8-bit quantization. This generated sample is not a photograph, an AITrending reproduction or evidence of Q3/Q4 quality. Source: MFlux contributors / MOSS-Image-SFT

MOSS-Image-SFT now has a complete native MLX implementation in MFlux, followed by ready-made community quantized packs listed at 5.8GB for Q3 and 7.5GB for Q4. The October 22 integration moves the Qwen3.5 text encoder into MLX too, making this more than another small image-transformer conversion for Apple Silicon.

MFlux pull request 914 merged at 00:03 UTC. It explicitly replaces proposal 895, the fork-based port covered on October 18. The earlier proposal was closed without merging; the replacement keeps the contributor’s work while completing the integration and addressing loading, framework and text-encoder issues.

The Q3 and Q4 repositories appeared around 00:53 UTC and date their conversions October 22. These are MFlux-Community builds of the existing OpenMOSS model, not a new OpenMOSS foundation-model release. Their repository inventories include the text encoder, tokenizer, transformer and VAE. The card’s download-size figures are not minimum unified-memory requirements.

A native text encoder changes the setup

The merged implementation runs the diffusion transformer, Qwen3.5 text encoder and VAE in MLX. Its maintainer says the earlier PyTorch text encoder and trust_remote_code=True path are gone. Quantization applies to the transformer and text encoder, and mflux-save can persist a converted copy.

The port supports natural-language, tag and JSON-layout prompts. Its documented defaults remain 50 steps and guidance 3; it is not the separate four-step MOSS turbo model. Dimensions must be multiples of 16, LoRA flags are unavailable for this model, and the scheduler-selection flag is ignored because the model uses its trained schedule.

Existing vanch007/MOSS-Image-SFT-mlx-* downloads are not interchangeable with this implementation. The new guide says those older packs have different weight names and omit the text encoder, so the new loader rejects them. Use the new community packages, the official upstream checkpoint or a copy produced by the current mflux-save workflow.

Use current source before trying the new command

There is a packaging gap at review. The community cards recommend uv tool install --upgrade mflux, but PyPI’s latest published 0.23.0 wheel was uploaded October 16. AITrending inspected that wheel’s file list and console entry points: it contains neither the MOSS model module nor mflux-generate-moss-image.

For the merged implementation, install the reviewed source revision rather than expecting that existing wheel to provide the command:

uv tool install --force "git+https://github.com/mflux-community/mflux.git@5e8644fdf2cd5838f80734e6b2822023f3f1c66c"

The new Q3 card documents this generation route:

mflux-generate-moss-image \
  --model mflux-community/moss-image-sft-mflux-q3 \
  --base-model moss-image-sft \
  --prompt "A wooden signpost in a forest reading HELLO WORLD" \
  --width 768 --height 768 --seed 42 --output image.png

The package targets Python 3.10 or later. Model files download on first use; the cards also document downloading them ahead of time and passing a local directory. This is local inference, not a paid hosted image API, though users still supply their hardware and download bandwidth. MFlux is MIT-licensed; the MOSS model cards declare Apache-2.0 licensing for the weights.

Read memory and quality numbers in context

The implementation notes report a 4-bit 1024-square run on an M-series Mac with 36GB, taking about 9.7 seconds per step and peaking at 10.6GB of MLX memory in the cited test. That is a maintainer measurement on one setup, not a Q3 memory guarantee or an AITrending benchmark.

The accompanying bicycle image is the project’s 640-by-368, 20-step, 8-bit regression reference, not an output generated by AITrending or evidence of Q3 quality. The news is a completed native pipeline and smaller complete distribution options, with a source-install requirement that readers should resolve before following the cards’ commands.