license: apache-2.0
library_name: comfyui
pipeline_tag: image-text-to-text
base_model: Aragduhl/Huihui-Qwen3-VL-8B-Instruct-abliterated-comfy
base_model_relation: quantized
tags:

• comfyui
• qwen3-vl
• nvfp4
• text-encoder
• blackwell
• abliterated
• uncensored

Huihui Qwen3-VL 8B Instruct Abliterated — ComfyUI NVFP4

Single-file ComfyUI comfy_quant NVFP4 conversion of
Aragduhl/Huihui-Qwen3-VL-8B-Instruct-abliterated-comfy,
which consolidates the full Qwen3-VL 8B abliterated checkpoint into one Transformers-format Safetensors file.

File

• Huihui-Qwen3-VL-8B-Instruct-abliterated-NVFP4.safetensors
• Size: 8,104,648,756 bytes
• SHA-256: 5c577e342acdaf869f4484fdf283336d69088c286cc5e800f1891f1c5f5b385f
• 1,464 stored tensors
• 238 text-transformer matrices quantized as NVFP4

Quantization policy

Conversion used convert-to-quant 1.3.3 with ComfyUI's comfy-kitchen 0.2.30 backend:

ctq \
  -i Huihui-Qwen3-VL-8B-Instruct-abliterated.safetensors \
  -o Huihui-Qwen3-VL-8B-Instruct-abliterated-NVFP4.safetensors \
  --nvfp4 \
  --qwen_vlm \
  --comfy_quant \
  --save-quant-metadata \
  --simple \
  --low-memory \
  --manual-seed 42

The qwen_vlm preservation policy keeps the complete visual encoder, embeddings, norms, biases, MTP-sensitive components, and the first and last language layers at source precision. The vision tower is not quantized. All 512 preserved tensors were compared against the source checkpoint bit-for-bit after conversion: no missing or mismatched tensors were found.

The resulting file contains native ComfyUI quantization metadata (_quantization_metadata, format version 1) and uses packed FP4 E2M1 weights with the scale tensors expected by comfy_kitchen.

Hardware

Native accelerated NVFP4 inference requires an NVIDIA Blackwell GPU and a current CUDA/PyTorch/ComfyUI stack. Conversion and validation were performed on:

• NVIDIA RTX PRO 6000 Blackwell Server Edition (SM120)
• PyTorch 2.13.0 + CUDA 13.0
• ComfyUI 0.31.0
• comfy-kitchen 0.2.30

On older GPUs the runtime may dequantize or fall back instead of using Blackwell FP4 Tensor Cores.

ComfyUI usage

Place the file in:

ComfyUI/models/text_encoders/

Load it with CLIPLoader:

• clip_name: Huihui-Qwen3-VL-8B-Instruct-abliterated-NVFP4.safetensors
• type: joyimage

It can then be connected to the native Generate Text / TextGenerate node or used in compatible workflows that expect the Qwen3-VL 8B text encoder.

Validation

The published artifact was validated in ComfyUI, not only by inspecting its header:

1. ComfyUI detected quantization metadata version 1.
2. CLIPLoader(type=joyimage) loaded the checkpoint successfully.
3. Native TextGenerate executed a real generation on the Blackwell GPU.
4. Test prompt: Reply with exactly: NVFP4 validation successful
5. Model output: NVFP4 validation successful

Credits and license

• Original Qwen3-VL model: Qwen team
• Abliterated checkpoint: huihui-ai
• ComfyUI single-file source: Aragduhl
• Quantization tooling: silveroxides/convert_to_quant and comfy-kitchen

License follows the Apache-2.0 source model.