Downloads · 30 days
103
2% of all-time downloads
fbjr/qwen3-vl-32b-W4A16-AWQ-H3
qwen3-vl-32b-W4A16-AWQ-H3 is a machine learning model from fbjr. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for minimax-h3. The card lists the license as apache-2.0.
Research / testing model. Published so others can try it and report back. It is not the most faithful option available — see below — and no blinded comparison of rendered output has been run.
Downloads · 30 days
103
2% of all-time downloads
All-time downloads
5.1K
Public
Repo size
40.8 GB
Likes
4
Public
Click a slice to open those files.
.safetensors40.8 GB · 100%
From the Hugging Face model README
Research / testing model. Published so others can try it and report back. It is not the most faithful option available — see below — and no blinded comparison of rendered output has been run.
note: there's probably no good reason to use this over the int8 convrot for daily use
A compressed-tensors W4A16 AWQ quantization of Qwen3-VL-32B, adapted for the 50-layer text/vision conditioner MiniMax H3 uses, plus the single-file ComfyUI node that loads it. ComfyUI supplies the H3 architecture and tokenizer; comfy-kitchen supplies the CUDA W4A16 operator. The node adapts compressed-tensors' namespace, packing and metadata in memory, because core does not recognize that format.
(width/32) * (height/32) tokens: 2048x2048 is 4,096, 512x512 is 256.longest_edge in the embedded
preprocessor_config.json near the top of
comfyui_minimax_h3_awq_loader.py. There is no widget for it.--fp16-text-enc, --fp32-text-enc and --fp8-*-text-enc do nothing here —
the compute dtype is fixed at bf16.
| file | what it is |
|---|---|
qwen3vl_32b_minimax_h3_w4a16_awq_v2-comfy.safetensors | the checkpoint |
comfyui_minimax_h3_awq_loader.py | the ComfyUI node |
comfyui_sample_workflows/ | sample workflows, from ComfyUI's H3 templates |
v2_hf_weights/ | the Hugging Face-format source, for anyone converting it themselves |
code/ | the quantization, conversion and validation scripts |
Checkpoint in models/text_encoders/, the .py at the top level of
custom_nodes/ (ComfyUI's scan only walks one level deep), workflows anywhere
under user/default/workflows/.
Median relative L2 at layer 50 against the BF16 reference in the same ComfyUI 50-layer implementation, over a 13-row holdout. Lower is closer.
| encoder | image refs, native | image refs, 2048 short edge | text only |
|---|---|---|---|
| int8_convrot (ComfyUI's own) | 0.0273 | 0.0212 | 0.0061 |
| nvfp4_awq | 0.3745 | 0.1472 | 0.0393 |
| this one | 0.3667 | 0.3326 | 0.0611 |
ComfyUI's INT8 ConvRot conditioner is roughly an order of magnitude closer to BF16. If layer-50 fidelity matters more than memory, use that instead. The reasons to use this one are size — about 19 GiB against about 25 — and that W4A16 executes on comfy-kitchen's CUDA path.
Layer-50 distance is a proxy. If you form an impression of the rendered output either way, that is more informative than this table.
This model uses the release's full still-image budget, so a reference prepared
at a 2048 short edge reaches the conditioner intact and costs
(width/32) * (height/32) merged tokens in the same sequence segment as your
prompt. Large references can crowd the prompt out of that segment, so if your
instructions bind identities to specific pictures, size the references
deliberately.