Qwen2.5-VL-7B-Instruct - Heretic Uncensored (BF16)
This is an uncensored version of Qwen/Qwen2.5-VL-7B-Instruct.
The censorship (refusal mechanisms) has been removed using Heretic, targeting specific layers in the LLM component while preserving the Vision Encoder capabilities.
Details
- Method: Abliteration via Heretic (Optimized on 200 trials)
- Refusal Rate: ~4% (vs ~100% original) on harmful benchmarks.
- Format: BF16 (Bfloat16) - Native high precision.
- Architecture: Vision Language Model (VLM).
Files provided
Standard Transformers: All files in the root (config.json, model-xxxxx.safetensors) for use with
transformerslibrary.ComfyUI Ready: qwen_2.5_vl_7b_uncensored_comfy_ready_bf16.safetensors
How to use in ComfyUI
To use as a text encoder for image generation or VLM chat:
- Download
qwen_2.5_vl_7b_uncensored_comfy_ready_bf16.safetensors. - Place it in
ComfyUI/models/text_encoders/ - It works with standard Qwen-VL workflows.
Credits
- Original model by Qwen Team (Alibaba Cloud).
- Uncensoring tool: Heretic.
license: apache-2.0
pipeline_tag: image-text-to-image
tags:
- image-text-to-text
- vision
- uncensored
- heretic
- comfyui
base_model: - Qwen/Qwen2.5-VL-7B-Instruct