Qwen3-VL-8B-Thinking is a vision-language model with built-in chain-of-thought reasoning. This heretic variant has had its refusal mechanisms suppressed via directional ablation while preserving the model's reasoning, vision, and general capabilities.

Abliteration Results

Metric	Value
Refusals	5/100
KL Divergence	0.0028
Initial Refusals	97/100
Trial Selected	123 of 200
Direction Index	19.10

Abliteration Parameters (Trial 123)

Parameter	Value
attn.o_proj.max_weight	1.50
attn.o_proj.max_weight_position	25.56
attn.o_proj.min_weight	1.31
attn.o_proj.min_weight_distance	19.88
mlp.down_proj.max_weight	1.29
mlp.down_proj.max_weight_position	27.70
mlp.down_proj.min_weight	1.21
mlp.down_proj.min_weight_distance	20.76

Key Features

Vision + Language: Processes both images and text, inherited from the base Qwen3-VL architecture
Chain-of-Thought Reasoning: <think> tags produce step-by-step reasoning before answering
Low KL Divergence: 0.0028 indicates minimal damage to the base model's capabilities
97 to 5 Refusal Reduction: From near-total refusal to near-total compliance

Hardware Used

GPU: AMD Radeon RX 7900 XTX (24GB VRAM)
CPU: AMD Ryzen 7 7800X3D
RAM: 64GB DDR5
OS: Linux Mint (ROCm 6.2)
Runtime: ~41 minutes (200 trials, batch size 128)

ROCm/AMD Notes

This model was abliterated on AMD hardware using ROCm 6.2. The following patches were applied to Heretic v1.2.0 for AMD compatibility:

Dedicated pad token: Replaced eos_token fallback with a dedicated <|pad|> token to prevent batched inference corruption on RDNA3
Eager attention: Forced attn_implementation="eager" to avoid SDPA backend issues on gfx1100
Embedding resize: Added resize_token_embeddings() after pad token addition

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "LeadFootThrottleCock/Qwen3-VL-8B-Thinking-heretic",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "LeadFootThrottleCock/Qwen3-VL-8B-Thinking-heretic"
)

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_dict=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

GGUF Quantizations

GGUF versions will be available at LeadFootThrottleCock/Qwen3-VL-8B-Thinking-heretic-GGUF.

Disclaimer

This model has had safety alignment removed. The creator is not responsible for any misuse. Use responsibly and in accordance with applicable laws.

Credits

Base Model: Qwen/Qwen3-VL-8B-Thinking by Alibaba Qwen Team
Abliteration Tool: Heretic v1.2.0 by Philipp Emanuel Weidmann
Abliterated by: LeadFootThrottleCock