DavidAU/Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-Thinking

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DavidAU/Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-Thinking is a 27 billion parameter Qwen 3.5-based causal language model, fine-tuned by DavidAU using Unsloth with custom datasets focused on Philip K. Dick's works. This model is uncensored and 'heretic-trained' to follow user instructions without refusal, and it features an upgraded Jinja template for improved tool handling and reduced repetition. It supports a 32,768 token context length and is optimized for creative writing, roleplay, and philosophical exploration in the style of Philip K. Dick, while also retaining multimodal vision capabilities.

Loading preview...

Model Overview

DavidAU/Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-Thinking is a 27 billion parameter causal language model based on the Qwen 3.5 architecture, fine-tuned by DavidAU. This model is specifically trained using Unsloth with five manually edited datasets centered around the works and themes of Philip K. Dick (PKD), aiming to deeply embed the 'PKD' persona without completely altering the base model's capabilities. It features an upgraded Jinja template to address issues like repetition and excessive 'thinking' in responses, and it includes enhanced tool handling.

Key Differentiators

  • PKD-Themed Fine-tuning: Specialized training on Philip K. Dick's works for generating content in his distinctive style and philosophical themes.
  • Uncensored & Heretic: Designed to be fully uncensored and to follow user instructions without refusal, trained post 'Heretic'ing.
  • Improved Template & Tooling: Incorporates an upgraded Jinja template to mitigate common LLM issues and enhance tool-use functionality.
  • Multimodal Capabilities: Retains and has been tested for vision (image) capabilities, with video portions noted as not tested.
  • Context Length: Supports a native context length of 32,768 tokens, extensible up to 1,010,000 tokens using YaRN scaling.

Performance & Usage

This model demonstrates competitive performance on various benchmarks, including language understanding, instruction following, and vision-language tasks, often outperforming the base Qwen3.5-27B model in specific areas. For instance, it shows improved scores in arc, arc/e, boolq, hswag, obkqa, piqa, and wino compared to Qwen3.5-27B-Text-VL qx86-hi. Notably, its refusal rate is significantly lower (14/100) compared to the original Qwen/Qwen3.5-27B (94/100), highlighting its uncensored nature. It is recommended for use with min q4ks or IQ3S quantization and a repetition penalty of 1 (off).

Ideal Use Cases

  • Creative Writing & Roleplay: Generating narratives, dialogues, or character interactions in the style of Philip K. Dick.
  • Philosophical Exploration: Discussing themes of reality, humanity, and artificial intelligence from a unique, uncensored perspective.
  • Unrestricted Content Generation: For applications requiring a model that adheres strictly to user prompts without safety alignment refusals.
  • Tool-Use Applications: Leveraging its enhanced tool-calling capabilities for agentic workflows, particularly with frameworks like Qwen-Agent and Qwen Code.

Finetunes
1 models