MMOSS Image AI

MOSS-Image vs FLUX.2 vs Z-Image

Three open-weight image model families people compare in October 2026. This page sticks to model cards and the MOSS-Image technical report. Where something is not published we write unknown. We have not run our own side-by-side benchmark.

Diagram of published LongLongTextBench scores
Diagram drawn from the authors' published LLTB table. Not our own benchmark and not a side-by-side photo.
Dark fantasy game-style scene, an OpenMOSS published output. OpenMOSS published MOSS-Image output, assets/images/011
Published output (assets/images/011). Not a FLUX or Z-Image result. Source: OpenMOSS assets/images/011. © OpenMOSS Team, Apache-2.0. We resized the long edge to at most 1400px and saved WebP. Not made on this site.

Spec sheet

MOSS-ImageFLUX.2 [dev]FLUX.2 [klein] 4BZ-Image / Turbo
DeveloperOpenMOSSBlack Forest LabsBlack Forest LabsTongyi-MAI (Alibaba)
On Hugging Face11 Oct 2026; announced 16 Oct 202622 Nov 202514 Jan 2026Turbo 25 Nov 2025; Z-Image 23 Jan 2026
Parameters5B DiT + frozen Qwen3.5-9BAbout 32.2BAbout 3.9B6B family (about 6.15B)
LicenseApache-2.0FLUX non-commercialApache-2.0Apache-2.0
Default steps50UnknownUnknownZ-Image 28–50; Turbo 8
Layout boxesYes, experimental JSONNot documented on the card we checkedUnknownNot documented on the card
Max prompt4K tokensUnknownUnknownUnknown
Stated VRAMNot publishedNot checkedNot checkedTurbo: fits within 16 GB (model card)

MOSS-Image uses the FLUX.2 VAE (loaded from FLUX.2-klein-4B), technical report Table 5. Parameter counts and dates were read from the Hugging Face API on 18 Oct 2026.

Long-text scores from the authors

LongLongTextBench (LLTB) is 240 tasks, 120 English and 120 Chinese, averaging about 1,340 characters per image. The MOSS-Image technical report Table 4 reports OCR coverage:

ModelEN longZH longMean of 6 groups
MOSS-Image97.192.496.3
Z-Image Base24.924.340.8
FLUX.24.82.911.8
GPT Image 2 (closed)97.398.299.0

LLTB was created and run by the MOSS-Image team, on their settings, for the task the model was designed for. The report does not name which FLUX.2 variant was tested. We have not reproduced these numbers. On the shorter LongTextBench the same report lists MOSS-Image at 97.63, close to several other models, which the authors call near-saturated.

Layout control

On 7Bench the report gives MOSS-Image mIoU 0.88 and AUC 0.70, versus 0.44 / 0.20 for the next model they tested (Ideogram 4.0). FLUX.2 and Z-Image were not in that table. These are the authors' evaluations.

Which one

  • Posters, menus, and documents with exact text, especially Chinese: MOSS-Image, with positions pinned in the layout builder.
  • Speed on a consumer GPU: Z-Image-Turbo's 8-step design and stated 16 GB fit. MOSS-Image defaults to 50 steps.
  • Commercial use: MOSS-Image, FLUX.2 [klein] 4B, and Z-Image are Apache-2.0. FLUX.2 [dev] is non-commercial.
  • Photorealism: we have no neutral benchmark to cite. Try the same prompt from the library on each model.

Sources include the technical report, the MOSS-Image-SFT card, FLUX.2-dev, FLUX.2-klein-4B, Z-Image-Turbo, and Z-Image.

Sources

Facts checked 2026-10-18.