DavidAU/Qwen3.5-21B-GLM-4.7-Flash-Deckard-Heretic-Uncensored-Thinking

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DavidAU/Qwen3.5-21B-GLM-4.7-Flash-Deckard-Heretic-Uncensored-Thinking is a 21 billion parameter language model, contracted from a 27B Qwen 3.5 base, and fine-tuned by DavidAU. It features a 32768 token context window and 48 layers, making it 33% faster and requiring less memory than its 27B counterpart. This model is explicitly uncensored, unfiltered, and designed for high reasoning, creative tasks, and complex instruction following, excelling in scenarios requiring character and intelligence.

Loading preview...

Overview

DavidAU/Qwen3.5-21B-GLM-4.7-Flash-Deckard-Heretic-Uncensored-Thinking is a 21 billion parameter language model, derived from a 27B Qwen 3.5 base. This model has been significantly optimized, featuring 48 layers and 639 Tensors, resulting in a 33% reduction in size and a 33% increase in speed compared to the base 27B model. It boasts a 256K context window and is specifically engineered for uncensored, unfiltered, and highly intelligent responses.

Key Capabilities

  • Uncensored and Unfiltered Output: Designed to provide responses without safety alignment or nanny-like restrictions, suitable for use cases requiring complete freedom of expression.
  • Enhanced Reasoning: Trained on the GLM 4.7 Flash High Reasoning dataset and internal "Deckard/PDK" datasets, it exhibits dramatically increased performance in reasoning and problem-solving.
  • Variable Length Reasoning: Adapts its reasoning depth, providing shorter responses for less complex queries and longer, more detailed explanations for intricate problems.
  • Creative Generation: Excels in creative writing tasks, especially when paired with specific system prompts and higher temperature/repetition penalty settings.
  • Code Generation & Image Processing: Tested and confirmed to handle code generation and image processing tasks effectively.
  • Optimized Performance: Achieves faster token generation and reduced memory footprint, making it efficient for deployment.

Good For

  • Creative Writing & Roleplay: Ideal for generating vivid, emotional, and detailed narratives, dialog, and character development, especially with tailored system prompts.
  • Unrestricted Content Generation: Suitable for applications where censorship or safety alignments are undesirable, allowing for raw and intense outputs.
  • Complex Instruction Following: Designed to carefully ponder and execute intricate user instructions, leveraging its advanced reasoning capabilities.
  • Resource-Efficient Deployment: Its contracted 21B parameter size and optimized architecture make it a faster and less memory-intensive option compared to larger models, while maintaining high intelligence.

Finetunes
1 models