nerkyor/Qwen3.8-27B-Uncensored-Heretic-T10-BF16

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

nerkyor/Qwen3.8-27B-Uncensored-Heretic-T10-BF16 is a 27 billion parameter Qwen3.8 derivative model, fine-tuned by nerkyor, specifically engineered to dissolve refusal tendencies. This model, with a 32768 token context length, focuses on reducing hard refusals to nearly zero, making it suitable for applications requiring less restrictive content generation. It achieves strong performance on benchmarks like MMLU500 (87.0%) and GPQA198 (85.35%), while significantly minimizing content refusal behaviors.

Loading preview...

Model Overview

This model, nerkyor/Qwen3.8-27B-Uncensored-Heretic-T10-BF16, is a 27 billion parameter derivative of the official Qwen3.8-27B base model, developed by nerkyor. Its primary distinction lies in its C0/T10 directional refusal-dissolve derivation, which significantly reduces the model's tendency to refuse prompts. This experimental release aims to study the technical feasibility and behavioral effects of refusal-tendency dissolution.

Key Capabilities & Features

  • Reduced Refusal Tendencies: Achieves 0.0% hard refusals on the Refusal140 benchmark for the BF16 version, and 0.7% for the Q8_0 version, making it highly permissive.
  • Strong General Performance: Demonstrates robust performance on reasoning and knowledge benchmarks, scoring 87.0% on MMLU500 and 85.35% on GPQA198.
  • Multimodal Support: Includes a Q8 vision projector, enabling multimodal capabilities when used with llama.cpp.
  • Optimized for llama.cpp: Provided with complete BF16 merge and verified Q8_0 llama.cpp packages, including an MTP draft for potential inference acceleration.
  • Configurable Reasoning Effort: Supports low, medium, and xhigh reasoning effort settings, allowing users to balance conciseness and depth in responses.

Ideal Use Cases

  • Research into AI Safety & Alignment: Particularly useful for studying the effects of refusal-tendency dissolution in large language models.
  • Applications Requiring Unrestricted Content Generation: Suitable for scenarios where strict content filtering or refusals are undesirable, provided users adhere to legal and ethical guidelines.
  • Exploration of Model Behavior: Developers and researchers can use this model to explore the boundaries of LLM responses and understand the impact of specific fine-tuning strategies on output behavior.

Finetunes
1 models