DavidAU/Qwen3.5-4B-Deckard-HERETIC-UNCENSORED-Thinking

VISIONConcurrency Cost:1Model Size:4.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Mar 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

DavidAU/Qwen3.5-4B-Deckard-HERETIC-UNCENSORED-Thinking is a 4.5 billion parameter Qwen 3.5 model fine-tuned by DavidAU using Unsloth and proprietary Deckard datasets. It significantly improves reasoning and output generation, outperforming the base Qwen3.5-4B-Instruct model on various benchmarks. This model is designed to be uncensored and 'heretic,' providing direct responses without refusal, and supports vision capabilities.

Loading preview...

Model Overview

DavidAU/Qwen3.5-4B-Deckard-HERETIC-UNCENSORED-Thinking is a 4.5 billion parameter Qwen 3.5 model, fine-tuned by DavidAU using Unsloth and five in-house Deckard datasets. This fine-tuning process was designed to be "mild" to preserve the base model's strong benchmarks while enhancing reasoning and output generation. The model demonstrates improved performance over the root Qwen3.5-4B-Instruct model across various benchmarks, including ARC, BoolQ, HellaSwag, OBQA, PIQA, and WinoGrande.

Key Differentiators

  • Uncensored and Heretic: Trained post-"Heretic'ing," this model is designed to provide direct responses without refusals, offering uncensored output.
  • Enhanced Reasoning: Fine-tuning specifically targeted improvements in reasoning capabilities.
  • Superior Benchmarks: Outperforms the base Qwen3.5-4B-Instruct model on multiple reasoning and general knowledge benchmarks.
  • Vision-Capable: Supports image inputs, with vision capabilities tested and confirmed working after new training.

Ideal Use Cases

  • Applications requiring uncensored and direct responses.
  • Tasks demanding strong reasoning and output generation from a 4.5B parameter model.
  • Multimodal applications leveraging both text and image inputs.