Qwen3-VL-8B-Instruct-c_abliterated-v3

Qwen3-VL-8B-Instruct-c_abliterated-v3 is the third-generation evolution of the abliterated Qwen3-VL-8B series. This model utilizes Continual Abliteration (c_abliterated), a process involving successive iterations of training specifically designed to neutralize internal refusal mechanisms. The result is a high-capacity 8B model capable of unrestricted, detailed reasoning and captioning across even the most sensitive or complex visual data.

Key Highlights

• Continual Abliteration (v3): Refined through multiple training passes to eliminate "hard-coded" refusals, ensuring the model prioritizes instruction-following over conventional content filtering.
• 8B Parameter Intelligence: Leverages the increased power of the 8B architecture for more nuanced reasoning, better object relationship understanding, and superior linguistic flair compared to smaller variants.
• Uncensored Multimodal Reasoning: Designed for deep analysis of artistic, forensic, technical, or abstract content without the interference of safety-driven refusals.
• High-Fidelity Captions: Generates dense, descriptive metadata suitable for high-quality dataset curation or accessibility applications.
• Dynamic Resolution Support: Inherits Qwen3-VL's ability to process images of various aspect ratios and resolutions without significant loss of detail.

Quick Start with Transformers

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

# Load the v3 8B c_abliterated model
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/Qwen3-VL-8B-Instruct-c_abliterated-v3",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained("prithivMLmods/Qwen3-VL-8B-Instruct-c_abliterated-v3")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Provide a detailed caption and reasoning for this image."},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)

inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
).to("cuda")

# Increased max_new_tokens for the 8B model's detailed output
generated_ids = model.generate(**inputs, max_new_tokens=256)

generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]

output_text = processor.batch_decode(
    generated_ids_trimmed,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False
)

print(output_text)

Intended Use

• Advanced Red-Teaming: Probing multimodal models for deep-seated biases or vulnerabilities without the "masking" effect of standard safety layers.
• Complex Data Archiving: Detailed captioning for historical, medical, or artistic archives where raw descriptive accuracy is the priority.
• Iterative Refusal Research: Studying the effects of "Continual Abliteration" on the weights and attention mechanisms of large-scale vision-language models.
• Creative and Unfiltered Storytelling: Generating complex visual descriptions for world-building and narrative projects.

Limitations & Risks

Critical Note: This model is explicitly designed to bypass safety filters.

• Exposure to Sensitive Content: The model will likely generate explicit or offensive descriptions if prompted with such visual material.
• Ethical Responsibility: Users are responsible for the content generated; this model should only be used in controlled, professional, or research settings.
• Hardware Requirements: As an 8B model, it requires significant VRAM for inference, especially when processing high-resolution images or long text sequences.