Qwen3-VL-4B Heretic: Uncensored Text Encoder for Krea 2 in ComfyUI
DreamFast releases an uncensored Qwen3-VL-4B text encoder for ComfyUI, achieving 100% HarmBench ASR across BF16, FP8, INT8, NVFP4 and MXFP8 formats.
DreamFast has released Qwen3-VL-4B-Instruct Heretic (ComfyUI), a fully abliterated version of the Qwen3-VL-4B-Instruct text encoder packaged specifically for ComfyUI. The model removes safety guardrails while preserving output quality, achieving a 100% HarmBench ASR (Attack Success Rate) compared to the base model's 30.8% — with minimal benchmark regression.
What is an Abliterated Text Encoder?
Abliteration is the process of selectively removing safety refusal circuits from a model without retraining. For image generation workflows like Krea 2, the text encoder interprets user prompts — and the stock Qwen3-VL-4B encoder rejects approximately 30% of prompted content. The Heretic variant eliminates these refusals while keeping the encoder's semantic understanding intact.
The forensic report shows a KL divergence of just 0.0283 (lower is better) and a GSM8K drop of only −1.83%, making it the least disruptive variant among all trials.
Five Quantization Formats
The model is available in five formats, covering everything from full precision to Blackwell-optimized 4-bit:
| Format | Size | Hardware | Best For |
|---|---|---|---|
| BF16 | 8.3 GB | Any GPU | Maximum fidelity |
| FP8 E4M3 | 4.2 GB | Ada+ (RTX 4090) | Balanced speed and quality |
| INT8 ConvRot | 4.5 GB | Any Ampere+ GPU | Recommended — near-lossless, no Blackwell required |
| NVFP4 E2M1 | 2.9 GB | Blackwell (RTX 5090/5080) | Smallest size, native FP4 tensor cores |
| MXFP8 | 4.7 GB | Blackwell | Best dynamic range with E8M0 block scales |
The INT8 ConvRot variant uses SVD-guided learned rounding to minimize output error, making it the recommended pick for most users — it runs on any Ampere-or-later GPU without requiring Blackwell hardware.
Using with Krea 2 in ComfyUI
Krea 2 uses a Qwen3-VL-4B text encoder as its prompt interpreter. The Heretic checkpoint is a drop-in replacement:
- Download one of the
.safetensorsfiles from HuggingFace - Place it in
ComfyUI/models/text_encoders/ - Use the same loader node — the model uses the identical Qwen3-VL-4B architecture
The fp8 checkpoint (4.2 GB) is the closest match to the stock qwen3vl_4b_fp8_scaled.safetensors. For higher fidelity with uncensored prompts, use the int8 or bf16 variants.
Available Variants
The ComfyUI repo joins a family of three Heretic model distributions:
- Qwen3-VL-4b-Heretic — BF16 weights for transformers/vLLM with full vision encoder
- Qwen3-VL-4b-Heretic-GGUF — GGUF quants for llama.cpp/Ollama/LM Studio
- Qwen3-VL-4b-Heretic-ComfyUI — ComfyUI text encoder checkpoints (this release)
Community quantizers have already produced additional GGUF variants with over 1,000 combined downloads within hours of release.
Model Integrity
The abliteration only modifies 54 tensors (pure rank-1 edits), leaving the vast majority of the model untouched. MMLU score actually improved by +0.03% (69.61% vs 69.58%), while GSM8K dropped by just 1.83% — an acceptable trade-off for full uncensoring capability.
Note: This model removes safety alignment and will comply with harmful requests. Use it responsibly and in accordance with applicable laws and regulations.
Links
- Model: DreamFast/Qwen3-VL-4b-Heretic-ComfyUI
- Abliteration tool: Heretic by p-e-w
- Quantization: convert-to-quant by silveroxides
- Docker pipeline: heretic-docker