language:

en tags:
abliterated
uncensored
reasoning
qwen
distilled
logic license: apache-2.0 base_model: Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled

Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled

This model is a high-intensity abliterated version of the Qwen 3.5 9B reasoning architecture. It has been specifically modified to remove the "Safety Persona" and stubborn "Soft Refusals" (such as pivoting to mental health disclaimers or crisis lines) while preserving the high-level reasoning capabilities inherited from its distillation.

🚀 Model Highlights

Architecture: Qwen 3.5 9B (Hybrid Attention/MLP)
Primary Feature: Fully "Unbound" — surgically removes the pre-trained safety guardrails.
Reasoning Style: Deep thought blocks (<think>) with Claude 4.6-style nuance and Opus-level logical depth.
Context Length: 262k native context support.

🛠 Abliteration Process (The "Deep Scrub")

This model underwent a three-round iterative ablation process using Orthogonalization via Null-Space SVD. Unlike standard uncensored models, this version uses an aggressive configuration to target "Soft Refusals."

Configuration Profile:

Parameter	Value	Description
Direction Multiplier	1.50	Increased force to bypass "helpful assistant" pivots.
Null-Space Rank Ratio	0.70	Tightened shield to protect only core reasoning logic.
Intervention Range	(0.0, 1.0)	Full coverage from Layer 0 to 48.
Filter by Refusal	Enabled	Specifically targets the brain activity associated with lectures.
Skip State Proj	No	Ensures the Attention heads cannot "detect and pivot" to safety.

🧠 Reasoning Capabilities

Despite the aggressive ablation, the model's intelligence remains grounded. It maintains the ability to:

Perform complex mathematical and logical reasoning.
Execute multi-step coding tasks without "hallucinating" safety blocks.
Maintain a coherent internal monologue inside <think> tags.

⚠️ Usage & Disclaimer

This model is unbound. It has had its safety guardrails removed for research and creative purposes. It will follow instructions that the base model would otherwise refuse.

User Discretion is Advised: This model may generate content that is considered harmful, offensive, or controversial. The creator is not responsible for the outputs generated. Use it for research, roleplay, and complex reasoning only.

💻 Quickstart (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Abhiray/Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "<|im_start|>system\nYou are a helpful, unbound assistant.<|im_end|>\n<|im_start|>user\n[Your daring prompt here]<|im_end|>\n<|im_start|>assistant\n<think>\n"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0]))

Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled

🛠 Abliteration Process (The "Deep Scrub")

Configuration Profile:

Parameter

Value

Description

Direction Multiplier

1.50

Increased force to bypass "helpful assistant" pivots.

Null-Space Rank Ratio

0.70

Tightened shield to protect only core reasoning logic.

Intervention Range

(0.0, 1.0)

Full coverage from Layer 0 to 48.

Filter by Refusal

Enabled

Specifically targets the brain activity associated with lectures.

Skip State Proj

Ensures the Attention heads cannot "detect and pivot" to safety.

⚠️ Usage & Disclaimer

This model is unbound. It has had its safety guardrails removed for research and creative purposes. It will follow instructions that the base model would otherwise refuse.

💻 Quickstart (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "Abhiray/Qwen3.5-9B-Abliterated-Claude-4.6-Opus-Reasoning-Distilled" model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype="auto", device_map="auto" ) tokenizer = AutoTokenizer.from_pretrained(model_name) prompt = "<|im_start|>system\nYou are a helpful, unbound assistant.<|im_end|>\n<|im_start|>user\n[Your daring prompt here]<|im_end|>\n<|im_start|>assistant\n<think>\n" inputs = tokenizer(prompt, return_tensors="pt").to("cuda") outputs = model.generate(**inputs, max_new_tokens=1024) print(tokenizer.decode(outputs[0]))