MOSS-Image
MOSS-Image (SFT) needs 7.3 GB of VRAM in ComfyUI, at the text encoder of its own template, by file sizes read 3 November 2026. The smallest card that loads it with room to work is the RTX 5070 (12 GB), and the template downloads 13.0 GB.
- Made by
- OpenMOSS
- Licence
- Apache 2.0
- Released
- October 2026
- Sizes read
- 3 November 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Which file on your card
For each memory size, the best file for that card: the highest-precision diffusion file that loads with room to work (one that only just fits when nothing else does), and what is left for working memory while it samples. On a larger card that can be above the file ComfyUI's template loads, which the answer above is about. "Streams" means even the smallest file is larger than the card: ComfyUI still runs it from system RAM, more slowly. FP4 files go only to the Blackwell cards that run them natively.
SFT
ComfyUI's template for SFT: INT8 ConvRot diffusion model, W4A8 text encoder. Stages: text encoder 7.3 GB, diffusion 5.7 GB. Download 13.0 GB.
Computed 3 November 2026Base
ComfyUI's template for Base: INT8 ConvRot diffusion model, W4A8 text encoder. Stages: text encoder 7.3 GB, diffusion 5.7 GB. Download 13.0 GB.
Computed 3 November 2026Check another card, or pick a file by hand, in the ComfyUI VRAM checker. Built your own workflow around it, with LoRAs or custom nodes? Drop it into the ComfyUI workflow checker.
Where to run it
SFT
On-demand prices before tax, read 19 October 2026.
Base
On-demand prices before tax, read 19 October 2026.
About the model
MOSS-Image is OpenMOSS's open image model. OpenMOSS's open text-to-image model from October 2026: a 5-billion-parameter single-stream transformer with a Qwen3.5 9B text encoder and the FLUX.2 small VAE. SFT is the instruction-tuned checkpoint ComfyUI's templates load; Base has no post-training or CFG distillation and is the one to train LoRAs on.
SFT: 5.1B · Base: 5.1B Text encoder: Qwen3.5 9B Apache 2.0 Released October 2026
Good at
- Text in images and planned layouts: it takes natural-language prompts, tag lists and JSON layouts, and ComfyUI has templates that place elements with boxes.
- Apache 2.0, for the weights and the code.
- Prompts of up to 4,000 tokens, and a prompt enhancer that runs on the text encoder already loaded, so there is no second language model to fit.
- LoRA training on small cards: KohakuMossTrainer documents adapter training on 8 GB and 6 GB cards.
Watch out for
- It has no image input, so it cannot copy a reference. The edit LoRA and the colour and depth control LoRA in Comfy-Org's repo come from community authors; check each one's own terms.
- ComfyUI's templates run 25 steps at CFG 2.5. OpenMOSS's own settings are 50 steps at CFG 3; try those when a result looks undercooked.
- In the layout templates, turning on the prompt enhancer can rewrite the text and drop the boxes.
- The templates need ComfyUI 0.42.0 or later.
Every MOSS-Image file ComfyUI loads
Sizes are the exact byte counts Hugging Face reports, read 3 November 2026. Files marked "template" are the ones ComfyUI's official workflow downloads.
| Part | Precision | Size | Version | File |
|---|---|---|---|---|
| Diffusion model | BF16 | 10.16 GB | SFT | moss_image_sft_bf16.safetensors |
| Diffusion model | INT8 ConvRot · template | 5.41 GB | SFT | moss_image_sft_int8_convrot.safetensors |
| Diffusion model | BF16 | 10.16 GB | Base | moss_image_base_bf16.safetensors |
| Diffusion model | INT8 ConvRot · template | 5.41 GB | Base | moss_image_base_int8_convrot.safetensors |
| Text encoder | BF16 | 19.31 GB | all | qwen3.5_9b_bf16.safetensors |
| Text encoder | INT8 ConvRot | 9.75 GB | all | qwen3.5_9b_int8_convrot.safetensors |
| Text encoder | W4A8 · template | 7.34 GB | all | qwen3.5_9b_w4a8_gptq.safetensors |
| VAE | BF16 · template | 0.25 GB | all | flux2-vae_small.safetensors |
What OpenMOSS says it needs
- "PyTorch GPU memory allocated / reserved | 3.65 / 3.85 GiB" (OpenMOSS's own pipeline, not ComfyUI, in its minimal profile, which keeps the transformer in system RAM: one 1024×1024 image at 50 steps, source)
Those figures and this page measure different things. A vendor's figure covers its own script at its default resolution, working memory included. This page counts the weights each stage holds, which is exact, and leaves the working memory as the space that remains, because it depends on your resolution and attention kernel and nobody publishes it per model. If generation runs out of memory with a file that loads, lower the resolution before stepping down a precision. ComfyUI's README says it "can run even the biggest open source models on as low as 4GB vram + 8GB ram" by streaming weights (source), so "streams" means slower, not impossible.
LoRA training for MOSS-Image
Only configurations a trainer's own documentation gives a VRAM figure for. Settings change the figure a great deal, so read the setting column before trusting the number.
| Trainer | VRAM | Setting | Source |
|---|---|---|---|
| KohakuMossTrainer | 6 GB | adapter training in its vram_6g layer: 2 resident blocks, activation and gradient offload, optimizer in system RAM, latents and text cached beforehand | docs |
Questions
- How much VRAM does MOSS-Image need?
- In ComfyUI's own workflow for SFT, the largest stage is the text encoder at 7.3 GB, and the whole workflow downloads 13.0 GB. Working memory for generation comes on top and grows with resolution. OpenMOSS itself says: "PyTorch GPU memory allocated / reserved | 3.65 / 3.85 GiB" (OpenMOSS's own pipeline, not ComfyUI, in its minimal profile, which keeps the transformer in system RAM: one 1024×1024 image at 50 steps).
- Can I run MOSS-Image on 8 GB of VRAM?
- Yes, with the INT8 ConvRot file: its largest stage, the text encoder, is 7.3 GB, which loads on 8 GB with 2.3 GB left for working memory, a tight fit.
- Can I run MOSS-Image on 12 GB of VRAM?
- Yes, with the INT8 ConvRot file: its largest stage, the text encoder, is 9.8 GB, which loads on 12 GB with 6.3 GB left for working memory.
- Can I run MOSS-Image on 16 GB of VRAM?
- Yes, with the BF16 file: its largest stage, the diffusion stage, is 10.4 GB, which loads on 16 GB with 5.6 GB left for working memory.
- Which MOSS-Image file should I download?
- The highest precision that loads fully on your card. On a 24 GB card that is the BF16 file (10.2 GB). ComfyUI's template downloads INT8 ConvRot with the W4A8 text encoder. FP4 files (NVFP4) are made for Blackwell cards (RTX 50, RTX PRO 6000, DGX Spark); on older cards, use FP8 or INT8.
- Files and sizes: Comfy-Org/MOSS-Image on Hugging Face, read from the Hugging Face API. Template files: ComfyUI's MOSS-Image guide.
- Model and licence: OpenMOSS-Team/MOSS-Image-SFT and the Apache 2.0.
- Card memory: manufacturers' specifications, as on each GPU page. The checker covers NVIDIA cards; ComfyUI also runs on AMD and Apple silicon, where file formats and speed differ and we have not sized them.
- Fit rule: a stage loads fully under 85% of usable memory and is tight up to 95%, the same margins the LLM pages use.
- Prices: each provider's own pricing page, on the dates beside them.