MOSS-Image vs Z-Image: which Apache 2.0 image model?
MOSS-Image draws long English and Chinese text and follows JSON layouts; Z-Image Turbo makes photoreal images in 8 steps on a 16 GB GPU. Both are Apache 2.0.
Verdict. Choose by what the picture must contain. MOSS-Image (a 5.08B generator, 27.2 GB with its Qwen3.5-9B text encoder) is built for dense, readable text and controllable layout: its report scores 96.3 on a long English and Chinese text benchmark and it follows JSON layouts that fix where each block of text goes, but it is text-to-image only and defaults to 50 steps. Z-Image (6.15B) is built for speed and photorealism: Turbo makes an image in 8 steps and fits a 16 GB GPU, and the base model is the established starting point for fine-tuning, though its editing and omni checkpoints are still unreleased. Both use plain Apache 2.0. MOSS-Image's report does not test Z-Image, so there is no published head-to-head on text or layout, and the best test is your own prompts.
Specifications
| Specifications | MOSS-Image | Z-Image |
|---|---|---|
| Developer | OpenMOSS | Tongyi MAI |
| Released | October 16, 2026 | November 26, 2025 |
| Total parameters | 5.08B (MOSS-Image-SFT (BF16)) | 6.15B (Z-Image-Turbo) |
| Active parameters | All | All |
| Architecture | Diffusion | Diffusion |
| Context window | – | – |
| Input | Text | Text |
| License | Apache 2.0 (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 4.2 GB (MOSS-Image-SFT (BF16)) | 16 GB (Z-Image-Turbo) |
Performance claims come from the developer and have not been independently verified.
Pick MOSS-Image if
- You make posters, infographics or notes with long English or Chinese passages that must be spelled correctly.
- You want to fix where each piece of text and each object goes with a JSON layout.
- You have a small GPU but plenty of system RAM and a fast SSD: the low-memory profile peaks near 4 GB of GPU memory and about 14 GiB of RAM.
- You want Base and SFT checkpoints for fine-tuning, with community trainers that list adapter training from 6 GB of VRAM.
Pick Z-Image if
- You want the fastest generation: Turbo needs 8 steps and no guidance, and fits a 16 GB GPU.
- You mostly make photorealistic images, which is where Tongyi says Turbo is strongest.
- You want the model with the larger community: Hugging Face showed about 743,000 downloads of Z-Image-Turbo in the 30 days to 11 November 2026, against about 22,000 for MOSS-Image-SFT.
- You can wait for Tongyi's editing and omni checkpoints, which its GitHub page lists as to be released.