license: "apache-2.0" pipeline_tag: image-text-to-text base_model:

• "Qwen/Qwen3.5-9B"
• "lukey03/Qwen3.5-9B-abliterated"
• "huihui-ai/Huihui-Qwen3.5-9B-abliterated" tags:
• qwen3.5
• vision-language
• multimodal
• abliterated
• int8
• convrot
• comfyui
• model-merging

Qwen3.5-9B 0% Refusal Vision — INT8 ConvRot

An experimental ComfyUI-oriented Qwen3.5-9B vision-language checkpoint. It combines the language weights from lukey03/Qwen3.5-9B-abliterated with the visual tower and visual merger weights from huihui-ai/Huihui-Qwen3.5-9B-abliterated. The language matrices were converted to INT8 ConvRot; the two already-quantized visual merger matrices retain the Huihui checkpoint's INT8 ConvRot weights and metadata.

The name “0% Refusal” refers to the upstream local checkpoint naming and is not a measured or guaranteed refusal rate. This is a community remix, not an official Qwen, lukey03, or huihui-ai release.

Lineage and attribution

This checkpoint is a merge of existing model weights, not a model trained from scratch:

• Base architecture and original model: Qwen/Qwen3.5-9B, by the Qwen team.
• Language model weights: lukey03/Qwen3.5-9B-abliterated. The local BF16 source checkpoint used for conversion had SHA-256 D739AFED0F1E4A05D1BAEC80FFBEEE624AB051AF857E95CBCEB189F20817C554, matching the Hugging Face LFS object hash for that repository's model.safetensors.
• Vision weights: huihui-ai/Huihui-Qwen3.5-9B-abliterated. Its 337 model.visual.* tensors were added to restore the vision tower and visual merger.
• Merge and ComfyUI ConvRot conversion: performed for this remix. No new pretraining, fine-tuning, or abliteration was performed as part of the merge itself.

The upstream model cards describe their own abliteration work and should be consulted for those methods. This project does not claim authorship of either upstream model or their abliteration procedures.

What is included

• One merged Qwen3.5-9B checkpoint: qwen3_abliterated-0%_vision_int8_convrot.safetensors.
• 1,264 tensors total, including 337 visual tensors.
• 250 language-model matrices stored as INT8 ConvRot, including the token embedding and language-model head.
• Two visual-merger matrices stored as INT8 ConvRot, with their scales and per-layer configuration.
• Remaining tensors retained in their source precision.
• ComfyUI quantization metadata for all 252 quantized matrices.

The checkpoint is approximately 9.83 GB (9.16 GiB). The INT8 ConvRot representation and metadata in this file target ComfyUI builds that support this quantization format; do not assume that the checkpoint loads directly through standard Transformers or other inference runtimes.

Use with ComfyUI

1. Use a ComfyUI version that supports Qwen3.5-9B text encoders and INT8 ConvRot mixed-precision weights.
2. Place the .safetensors file in ComfyUI/models/text_encoders/.
3. Restart or refresh ComfyUI's model list.
4. Select it with a Qwen3.5-compatible text-encoder or multimodal node that accepts image inputs.

Exact node names and image-input support depend on the ComfyUI version and installed custom nodes.

Validation

The merged checkpoint was verified locally:

• ComfyUI identifies it as QWEN35_9B.
• The Qwen3.5 vision module loads all 337 visual tensors with no missing or unexpected keys.
• A synthetic image-patch forward pass through the complete vision tower and merger produced finite output with shape (4, 4096).
• The user also confirmed that the merged checkpoint works in their ComfyUI workflow.
• For the 250 language matrices converted from BF16, the measured relative L2 quantization error averaged 0.9713%, with a maximum of 1.0652%. This measures conversion error against the BF16 source; it is not a behavioral or safety benchmark.

These checks do not establish general vision-language benchmark performance, a particular refusal rate, or compatibility with every ComfyUI workflow.

Limitations and responsible use

This is an experimental model merge. Reduced refusal behavior is not the same as accuracy, reliability, or permission to use the model for every purpose. Outputs can be incorrect, biased, sensitive, or inappropriate. Review outputs and follow applicable laws, platform rules, and the licenses of the upstream models.

The “0% Refusal” wording is descriptive of the upstream variant's naming, not a guarantee that every refusal has been removed or that the model will answer every request. Vision behavior after combining language and vision components may differ from either upstream model and has not been comprehensively benchmarked here.

Licenses

The upstream model cards identify the listed Qwen, lukey03, and huihui-ai checkpoints as Apache-2.0. This repository is marked Apache-2.0 based on those upstream declarations. Include the applicable Apache-2.0 license text and preserve required notices when redistributing the weights. Review the upstream repositories and their license files before redistribution; this card is not legal advice.

References

• Qwen team, Qwen3.5-9B
• lukey03, Qwen3.5-9B-abliterated
• huihui-ai, Huihui-Qwen3.5-9B-abliterated
• Arditi et al., Refusal in Language Models Is Mediated by a Single Direction
• Sumandora, remove-refusals-with-transformers