MOSS-Image
A 5B single-stream transformer in the FLUX.2 latent space, paired with a Qwen3.5 text encoder that is larger than the image model itself. Apache 2.0 for weights and code, with strong English and Chinese text rendering and JSON layout control.
What you actually download
A diffusion pipeline is three separate models. These sizes are read from the real .safetensors files in the Hugging Face repository, not estimated.
Peak VRAM at 1024×1024, one image
Every combination of precision and offload strategy. Lower precision costs image quality; offloading costs speed.
| Precision | Everything in VRAM | Text encoder on CPU | Sequential CPU offload |
|---|---|---|---|
| BF16 / FP16Reference quality. What the repos ship by default. | 30.0 GB | 13.2 GB | 19.6 GB |
| FP8Roughly half the weight memory. Small quality cost, widely used for Flux and SD 3.5. | 16.5 GB | 8.1 GB | 11.2 GB |
| NF4 / INT4Aggressive 4-bit. Visible quality loss, but it puts big models on small cards. | 14.5 GB | 6.1 GB | 11.2 GB |
Which machines can run it
75 of the 79 devices we profile can run MOSS-Image at 1024×1024 in some configuration. Showing the best-quality setup each one supports.
Showing the 30 devices with the most memory. The full hardware matrix covers all 79.
Licence: Apache 2.0
Yes. MOSS-Image is released under Apache 2.0, which permits commercial use. Keep the licence and attribution notices with any redistribution.
Compare model licences →MOSS-Image: common questions
- How much VRAM does MOSS-Image need?
- At 1024x1024 and batch 1, MOSS-Image peaks at about 30.0GB with everything resident at BF16, and as little as 6.1GB at NF4 + text encoder on CPU. The weights alone are 27.2GB at BF16 — 10.2GB of transformer, 16.8GB of text encoder and 0.25GB of VAE.
- Which GPUs can run MOSS-Image?
- 75 of the 79 devices we profile can run it at 1024x1024 in some configuration. The Apple Mac Studio (M5 Ultra, 512GB, 2026) handles it best, at about 30.0GB. Lower precision and text-encoder offload widen that list considerably.
- Can MOSS-Image be used commercially?
- Yes. MOSS-Image is released under Apache 2.0, which permits commercial use. Keep the licence and attribution notices with any redistribution.
- Does offloading the text encoder help on MOSS-Image?
- Yes, substantially. MOSS-Image uses a large language model as its text encoder — 16.8GB at full precision. Encoding happens once per prompt, so moving it to CPU costs a couple of seconds and frees that 16.8GB for the rest of the run, taking peak VRAM at BF16 from 30.0GB down to 13.2GB.
Weight sizes read from the model repository on 2026-09-12. Activation memory is an approximation, not a benchmark — BFWR does not run a hardware test lab.
Other open image models
Guides that go deeper
Setting this model up in ComfyUI, its licence, and the hardware it needs.
- MOSS-Image in ComfyUI: Native Nodes, Files, Memory and SettingsThe Apache 2.0 model, the nightly-only native nodes and the noise trap.Read the guide
- Qwen-Image 2.1 in ComfyUI: Files, VRAM, Turbo and the LicenceWhat to download, the memory it needs, and its research-only licence.Read the guide
- How to Run Stable Diffusion Locally: Hardware, VRAM and SetupCurrent models from SD 1.5 to Qwen-Image, and the VRAM each needs.Read the guide