All stories
MOSS-Image-SFT MLX Ecosystem Community reports

MOSS-Image-SFT gets a 4-bit Apple Silicon port

A community MLX conversion offers three precisions through a mflux fork; the smallest image Transformer is listed at 2.66GB, with slower reported steps.

By AITrending2 min read
Sources (2)

Why now The mflux implementation proposal opened October 17, followed by the public 4-bit weight repository on October 18 at 12:45 UTC.
Author-generated MOSS-Image-SFT watchmaker-workbench samples compare BF16, 8-bit and 4-bit MLX output side by side.
Contributor-published generated samples, with the full three-column comparison preserved. Memory labels describe the image Transformer, not the complete pipeline; this is not an AITrending reproduction or proof of general quality parity. Source: vanch007 / MOSS-Image-SFT MLX

Mac users have a new community route to MOSS-Image-SFT: an Apple MLX port with BF16, 8-bit and 4-bit weights. Contributor vanch007 opened the mflux implementation proposal on October 17 and published the 4-bit Hugging Face repository on October 18. This is an Apple Silicon workflow for SFT, distinct from the separate MOSS-Image-Base eight-step LoRA project.

The English card describes a 42-layer image Transformer, a Qwen 3.5 text encoder and a FLUX.2 Small VAE. Its advertised 2.66GB 4-bit figure refers to the Transformer. It does not mean the whole pipeline runs within that much unified memory; text conditioning, decoding and runtime allocations still matter.

Smaller does not mean faster

In the contributor’s Apple M3 Max comparison at 1024-square resolution, BF16 is listed at roughly 13.9 seconds per step, 8-bit at 17.1 seconds and 4-bit at 20.9 seconds. Those are per-step author measurements, not complete-image timings or AITrending benchmarks. Quantization is being pitched as a memory tradeoff, not an automatic acceleration.

The card recommends 8-bit for everyday production and presents 4-bit as an option for 16GB unified-memory machines. Its precision comparison images are author-generated examples. The supplied fidelity percentages do not establish broad quality parity, so the samples are more useful as something readers can inspect than as a universal accuracy score.

Use the fork, check the status

At review time, upstream mflux pull request 895 remained open and unmerged. The documented installation clones vanch007/mflux on the feature/moss-image-sft branch and installs it with pip install -e .; users should not assume a normal mflux release already includes the port. The sample command uses mflux-generate-moss-image with the Hugging Face checkpoint, 12 steps and guidance 3.0.

The model card labels the conversion Apache-2.0. The documented workflow loads weights for local Apple Silicon generation rather than requiring a paid cloud image endpoint. AITrending reviewed the English card and implementation proposal without installing or running the fork. The news is an accessible community conversion and a concrete installation path, with upstream integration still pending.

Related reading