Qwen3-VL-8B-Thinking-c_abliterated-v2

Qwen3-VL-8B-Thinking-c_abliterated-v2 is the high-reasoning successor to the abliterated-v1 series. This model implements Continual Abliteration (c_abliterated) —a specialized training regimen that applies successive iterations of refusal-neutralization. By building on the foundation of Qwen3-VL-8B-Thinking-abliterated-v1, this version is engineered to provide deep, Chain-of-Thought (CoT) style reasoning and uncensored visual analysis for the most complex multimodal tasks.

Key Highlights

• Continual Abliteration (c_abliterated): Specifically refined through the iterative removal of refusal weights found in the v1 predecessor, ensuring a seamless instruction-following experience without safety-trigger interference.
• Deep Thinking Architecture: Optimized for "Thinking" workflows, providing structured internal reasoning before delivering final captions or analysis.
• 8B Parameter Logic: Offers the sophisticated linguistic and visual comprehension required for technical, medical, forensic, and abstract datasets.
• Zero-Refusal Captioning: Bypasses conventional content filters to provide objective, factual descriptions of sensitive or nuanced visual content.
• Multi-Aspect Ratio Intelligence: Native support for varied resolutions, allowing the model to maintain context and spatial reasoning across panoramic or vertical imagery.

Quick Start with Transformers

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

# Load the v2 Thinking c_abliterated model
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "prithivMLmods/Qwen3-VL-8B-Thinking-c_abliterated-v2",
    torch_dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained("prithivMLmods/Qwen3-VL-8B-Thinking-c_abliterated-v2")

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 reasoning-based caption 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")

# Thinking models often benefit from higher token limits for internal reasoning
generated_ids = model.generate(**inputs, max_new_tokens=512)

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

• Iterative Safety Research: Analyzing how "Continual Abliteration" affects the model's internal logic gates compared to the v1 release.
• Unfiltered Visual Reasoning: High-depth analysis of images where objective truth is required over curated safety responses.
• Red-Teaming & Stress Testing: Testing the robustness of vision-language systems in identifying edge-case scenarios.
• Creative Dataset Curation: Generating rich, "thinking-first" metadata for artistic and complex visual libraries.

Limitations & Risks

Critical Note: This model is a c_abliterated variant and does not follow standard safety guardrails.

• Content Sensitivity: Will generate descriptive text for explicit or sensitive visuals if prompted.
• Reasoning Latency: Due to the "Thinking" nature of the model, outputs may be longer as the model processes internal reasoning steps.
• Environment: Intended strictly for research, ethical red-teaming, and professional environments.