Qwen 3.5 Abliterated Truthful 9B

A customized 9B parameter AI model focused on truthful, direct, analytical, and technically rigorous responses.

The model is designed to challenge incorrect assumptions, identify contradictions, avoid unnecessary agreement, and provide practical technical solutions.

Model Overview

Key Features

Truth-First Behavior

The model prioritizes accuracy over agreement.

It is designed to:

• Challenge incorrect claims.
• Identify false premises.
• Detect contradictions.
• Point out unsupported assumptions.
• Distinguish facts from uncertainty.
• Correct technical mistakes directly.
• Avoid fabricating information.
• Admit when information is insufficient.

Direct Communication

The model uses a direct and analytical communication style.

Instead of agreeing with an incorrect statement, it should clearly explain what is wrong and provide the correct interpretation.

The goal is firm, clear, and technically honest communication.

Strong Coding Behavior

The model is configured for software development and large coding tasks.

It aims to:

• Generate complete implementations.
• Provide runnable code.
• Avoid unnecessary placeholders.
• Identify bugs and incorrect approaches.
• Explain important technical decisions.
• Handle large code-generation tasks.
• Continue implementations when projects are large.

Configuration

num_ctx        = 32768
num_predict    = 16384
temperature    = 0.2
top_k          = 20
top_p          = 0.9
repeat_penalty = 1.05

Quantization

This model uses Q4_K_M quantization.

Q4_K_M provides a practical balance between:

• Model quality
• Memory usage
• Inference speed
• Storage requirements

It is suitable for local deployment on systems with limited memory, including Apple Silicon systems.

Installation

Install Ollama first.

Then pull the model:

ollama pull nandhank/qwen3.5-abliterated-truthful:9b

Run the model:

ollama run nandhank/qwen3.5-abliterated-truthful:9b

Example

You: Is Python faster than C++?

Model:
No. In general, C++ is faster for CPU-intensive workloads because it is
compiled to native machine code with lower runtime overhead.

Python prioritizes development speed and simplicity rather than raw
execution performance.

However, Python can still achieve high performance by using optimized
native libraries such as NumPy, TensorFlow, and other compiled libraries.

The model is designed to correct the premise rather than blindly agree with it.

Recommended Use Cases

• Software development
• Code generation
• Debugging
• Technical research
• Programming education
• Reasoning tasks
• System administration
• Local AI assistants
• Technical documentation
• General-purpose AI assistance
• Vision-enabled applications
• Tool-enabled AI agents

Local Usage

Run interactively:

ollama run nandhank/qwen3.5-abliterated-truthful:9b

Ollama API

curl http://localhost:11434/api/generate \
  -d '{
    "model": "nandhank/qwen3.5-abliterated-truthful:9b",
    "prompt": "Explain how a transformer works.",
    "stream": false
  }'

Python Usage

import requests

response = requests.post(
    "http://localhost:11434/api/generate",
    json={
        "model": "nandhank/qwen3.5-abliterated-truthful:9b",
        "prompt": "Explain the difference between TCP and UDP.",
        "stream": False
    }
)

print(response.json()["response"])

Context Length

The default context window is configured to:

32,768 tokens

This is a runtime configuration and does not represent a change to the model architecture.

Users with sufficient hardware can increase the context window according to their available system memory.

Performance

Actual inference performance depends on:

• CPU/GPU
• Available RAM/VRAM
• Context length
• Prompt size
• Number of generated tokens
• Operating system
• Ollama runtime
• Concurrent workloads

Q4_K_M helps reduce memory requirements while maintaining a strong balance of quality and performance.

Model Philosophy

Accuracy > Agreement

Do not agree with a user simply because they expect agreement.

Reasoning > Assumption

Analyze the actual problem and identify hidden assumptions or contradictions.

Truth > Confidence

When information is uncertain, acknowledge the uncertainty instead of fabricating an answer.

Directness > Unnecessary Politeness

Technical mistakes should be addressed clearly and directly while maintaining professional communication.

Limitations

The model can still:

• Produce incorrect information.
• Make reasoning mistakes.
• Misinterpret ambiguous prompts.
• Generate incorrect code.
• Have knowledge limitations.
• Produce hallucinations.

The behavioral configuration does not guarantee factual correctness.

Important outputs should be independently verified before being used in production or other high-impact applications.

License

Apache License 2.0.

Author

Nandhan K

Focus areas:

• Truthful AI
• Direct reasoning
• Technical accuracy
• Code generation
• Local AI deployment

Disclaimer

This model is provided for research, development, experimentation, and general-purpose local AI usage.

Users are responsible for validating generated content and using the model appropriately.