MOSS-Image HMS v1
MOSS-Image in 3.96 GB, about 2x faster, with the stock ComfyUI loader.
HMS v1 is the HaDeS MOSS Suite's own quantization of MOSS-Image. It is a mix of ComfyUI's native W6A8 and W4A8 formats, calibrated on the model itself with GPTQ.
There is no custom loader and no custom node to run it. Put it in models/diffusion_models and load it with Load Diffusion Model.
More info: https://huggingface.co/DeVilDoNia/MOSS-Image-HMS
Performance comparison
MetricBF16int8 ConvRot (Comfy-Org)HMS v1File size10.2 GB5.4 GB3.96 GBPeak VRAM (RTX 5090, 1024²)17.8 GB—11.6 GBSpeed vs BF161x—about 2x

Measured against BF16
RTX 5090, same seeds and settings.
ModelDistance from BF16 (L1@64, 10 scenes)Brightness shift (median luma)Text read exactly (HMS / BF16)Base4.13%+0.49 / 255139/143 vs 140/143SFT4.72%+0.70 / 255106/110 vs 105/110
The lettering reads as well as BF16, and the brightness does not drift. In a blind review against BF16, with the keys unopened, the four HMS v1 models we made were judged practically equal.

What's inside
Attention in all 42 blocks: W6A8.
FFN of blocks 30–41: W6A8. Measured against the other layouts, these blocks were where 6 bits paid off.
FFN of blocks 0–29: W4A8 with a 4-bit codebook.
Everything else: (embedders, modulation, norms, final layers): bfloat16.
The calibration runs on seven fixed prompts sampled with the model itself: four scenes and three with quoted text. It sets the scales and codebook by real activation energy, then rounds each projection with GPTQ in the rotated basis.
Leaving out the text prompts cost lettering (Base 131/143 without them, 139/143 with them), so they stay.
The full experiment, including the rejected candidates, is in the Suite's tools/RESULTS.md on GitHub.
How to use
1. Diffusion model
Place MOSS-Image_SFT_HMS_v1.safetensors or MOSS-Image_Base_HMS_v1.safetensors in ComfyUI/models/diffusion_models/.
2. Text encoder
Load CLIP, type moss_image:
qwen3.5_9b_int8_convrotqwen3.5_9b_w4a8_gptq
Both from Comfy-Org/MOSS-Image.
3. VAE
flux2-vae
4. Generation settings
The same as BF16. Our figures used:
ParameterValueResolution1024 × 1024Steps25CFG2.5Samplerer_sdeSchedulersimpleShift5
Requirement: a recent ComfyUI with comfy-kitchen's W4A8 / W6A8 kernels (tested with comfy-kitchen 0.2.40, RTX 5090). Speed and memory were measured only on that GPU.
Recommended: HaDeS MOSS Suite
These files were made with, and work best inside, the HaDeS MOSS Suite: free, Apache 2.0.
Comfy Registry:
hades-moss-suiteVideo tutorials: https://youtube.com/@Devildonia
MOSS Loader
Model, text encoder and VAE in one node. It loads HMS files as they are.
MOSS Model Saver (v0.9.5+)
Saves any MOSS-Image model in HMS v1, merges and baked LoRAs included.
Set
precisiontoHMS v1 (W4A8 / W6A8, calibrated).Connect the text encoder to
clip.Set the calibration steps and CFG the model is meant to run with.
For SFT the result is bit-identical to this file (942/942 tensors).
A save takes 8–10 min on an RTX 5090 (peak about 19.5 GB RAM, 24.6 GB VRAM).
A merge in HMS can come out slightly darker than its BF16 save (median luma −2.8 on our test merge), so check a merge by eye.
MOSS Compare Sampler / Compare Sheet
Put BF16 and HMS side by side with the same seed and see the difference yourself.
MOSS Text Check
Reads the text in an image and checks it against the prompt. It is the tool we used to score lettering here.
Limits
Not identical: HMS v1 is close to BF16 but not the same. With the same seed small details change, and a busy layout (a manga page, for example) can redraw one panel differently.
Training: to train or merge, start from the BF16 model. Requantizing an HMS file is not supported.
Credits
Source models: MOSS-Image Base and SFT, by the OpenMOSS Team (Apache 2.0).
BF16 weights: from Comfy-Org/MOSS-Image.
Formats and kernels: ComfyUI / comfy-kitchen.
HMS v1 recipe and files: HaDeS (DeVilDoNia).

