DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED model is a 9 billion parameter instruction-tuned variant of the Qwen 3.5 dense model, fine-tuned by DavidAU using a Claude 4.6 distill dataset. This model is designed to be uncensored and follows instructions without refusal, making it suitable for applications requiring direct and unfiltered responses. It also supports vision capabilities and has a native context length of 32768 tokens, extensible up to 1,010,000 tokens with YaRN scaling.

Loading preview...

Model Overview

This model, DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT-HERETIC-UNCENSORED, is a 9 billion parameter instruction-tuned variant of the Qwen 3.5 dense model. It was fine-tuned by DavidAU using a large Claude 4.6 distill dataset, with a focus on maintaining the strong benchmarks of the original Qwen 3.5 model while introducing specific modifications.

Key Differentiators

  • "Heretic" and Uncensored: This model is explicitly trained to follow instructions "no questions asked" and is fully uncensored, offering direct responses without refusal. This is a significant departure from models with built-in safety or refusal mechanisms, as evidenced by its 6/100 refusal rate compared to the original model's 100/100.
  • Instruction-Tuned: It operates in an "instruct only" mode, utilizing a modified Jinja template for instruction following.
  • Multimodal Capabilities: The model supports vision (image) inputs, with video understanding also noted as a feature, though not fully tested in this specific release.
  • Extended Context Length: Natively supports a context length of 262,144 tokens, and can be extended up to 1,010,000 tokens using YaRN scaling techniques.

Performance Highlights

Compared to the base Qwen3.5-9B model, this fine-tuned version shows improved performance on several benchmarks:

  • ARC: 0.574 (vs 0.417 for base Qwen3.5-9B)
  • HSwag: 0.714 (vs 0.634 for base Qwen3.5-9B)
  • PIQA: 0.780 (vs 0.737 for base Qwen3.5-9B)

Use Cases

This model is particularly suited for:

  • Applications requiring unfiltered responses: Due to its uncensored nature and "heretic" training, it will execute commands directly without moral or ethical filtering.
  • Instruction-following tasks: Optimized for direct instruction execution.
  • Multimodal tasks: Capable of processing image inputs, with potential for video understanding.
  • Long-context processing: Ideal for tasks requiring analysis or generation over very long texts, thanks to its extended context window.

Finetunes
2 models