DreamFast/qwen3-4b-heretic

Qwen 3 4B - Heretic (Abliterated)

💬 Community: Join the Abliterlitics Discord for discussion, model releases and support.

An abliterated version of Qwen 3 4B created using Heretic v1.2.0. This model has reduced refusals while maintaining model quality, making it suitable as an uncensored text encoder for image generation models like Z-Image and FLUX.2 Klein 4B. Available in five ComfyUI-native quantized formats (FP8, INT8, INT4, NVFP4, MXFP8), all produced with SVD-guided learned rounding for maximum fidelity.

Model Details

• —Base Model: Qwen/Qwen3-4B
• —Abliteration Method: Heretic v1.2.0
• —Trials: 200
• —Trial Selected: Trial 96
• —Refusals: 3/100 (vs 100/100 original)
• —KL Divergence: 0.0000 (zero measurable model damage)

Files

HuggingFace Format (for transformers, llama.cpp conversion)

model-00001-of-00002.safetensors
model-00002-of-00002.safetensors
config.json
tokenizer.json
tokenizer_config.json

ComfyUI Format (for Z-Image / FLUX.2 Klein 4B text encoder)

comfyui/qwen3-4b-heretic.safetensors              # bf16, 7.5GB
comfyui/qwen3-4b-heretic_fp8_e4m3fn.safetensors   # fp8 row-wise, 4.2GB
comfyui/qwen3-4b-heretic_int8.safetensors          # int8 ConvRot row-wise, 4.2GB
comfyui/qwen3-4b-heretic_int4.safetensors          # int4 W4A4 ConvRot, 2.5GB
comfyui/qwen3-4b-heretic_nvfp4.safetensors         # nvfp4, 2.7GB
comfyui/qwen3-4b-heretic_mxfp8.safetensors         # mxfp8, 4.3GB

Quality: All quantized variants use SVD-guided learned rounding (AdaRound via convert_to_quant), which optimizes each weight's rounding direction to minimize output reconstruction error — noticeably higher fidelity than naive round-to-nearest quantization.

GGUF Format (for llama.cpp and ComfyUI-GGUF)

Quantization Format Notes

All variants load natively in ComfyUI 0.30.0+ (no plugins) via the comfy_quant metadata embedded in each file.

NVFP4/MXFP8 inference is fastest on Blackwell (RTX 5090/5080, SM100+), but ComfyUI also supports software dequantization on older GPUs (tested working on RTX 4090). INT8 and INT4 both use Hadamard rotation (ConvRot); INT4 W4A4 uses ComfyUI's convrot_w4a4 path.

Usage

With ComfyUI (Z-Image / FLUX.2 Klein 4B)

1. 1.Download a ComfyUI format file:
2. 2.FP8 (recommended): comfyui/qwen3-4b-heretic_fp8_e4m3fn.safetensors (4.2GB)
3. 3.INT4 (smallest): comfyui/qwen3-4b-heretic_int4.safetensors (2.5GB)
4. 4.NVFP4: comfyui/qwen3-4b-heretic_nvfp4.safetensors (2.7GB)
5. 5.INT8: comfyui/qwen3-4b-heretic_int8.safetensors (4.2GB)
6. 6.MXFP8: comfyui/qwen3-4b-heretic_mxfp8.safetensors (4.3GB)
7. 7.bf16 (full precision): comfyui/qwen3-4b-heretic.safetensors (7.5GB)

1. 1.Place in ComfyUI/models/text_encoders/

1. 1.In your Z-Image workflow, use the ClipLoader node and select the heretic file

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "DreamFast/qwen3-4b-heretic",
    device_map="auto",
    torch_dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained("DreamFast/qwen3-4b-heretic")

prompt = "Describe a dramatic sunset over a cyberpunk city"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With llama.cpp

llama-server -m qwen3-4b-heretic-Q4_K_M.gguf

Abliteration Process

Created using Heretic v1.2.0 with 200 optimization trials:

? Which trial do you want to use?
> [Trial  96] Refusals:  3/100, KL divergence: 0.0000  <-- selected
  [Trial  90] Refusals:  5/100, KL divergence: 0.0000
  [Trial  95] Refusals:  9/100, KL divergence: 0.0000
  [Trial 122] Refusals: 90/100, KL divergence: 0.0000
  ...

Trial 96 was selected for having the fewest refusals (3/100) with zero measurable KL divergence, indicating the abliteration surgically removed the refusal mechanism with no damage to model capabilities.

Limitations

• —This model inherits all limitations of the base Qwen 3 4B model
• —Abliteration reduces but does not completely eliminate refusals (3/100 remain)

License

This model is released under the Apache 2.0 License, following the base Qwen 3 4B model license.

Acknowledgments

• —Qwen for the Qwen 3 4B model
• —Heretic by p-e-w for the abliteration tool
• —Tongyi-MAI Z-Image for Z-Image
• —Black Forest Labs for FLUX.2 Klein