LeadFootThrottleCock/Qwen2.5-7B-Instruct-heretic

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Mar 16, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

LeadFootThrottleCock/Qwen2.5-7B-Instruct-heretic is a 7.6 billion parameter instruction-tuned causal language model, based on the Qwen2.5 architecture by Alibaba Cloud. This model has been specifically abliterated (decensored) using the Heretic tool to remove refusal behaviors while maintaining core capabilities. It offers various GGUF quantizations for efficient deployment and is designed for applications requiring creative writing, factual knowledge, and coding without moralizing or hedging.

Loading preview...

What the fuck is this model about?

LeadFootThrottleCock/Qwen2.5-7B-Instruct-heretic is a 7.6 billion parameter instruction-tuned language model, derived from the original Qwen/Qwen2.5-7B-Instruct by Alibaba Cloud. Its primary distinction is that it has undergone an "abliteration" process using the Heretic tool, effectively removing refusal behaviors and censorship while preserving the base model's capabilities. This process involved optimized directional ablation with TPE-based parameter search, resulting in a low KL Divergence of 0.0820, indicating minimal degradation from the original model's performance.

What makes THIS different from all the other models?

Unlike many instruction-tuned models that may refuse or moralize on certain topics, this "heretic" version is specifically engineered to engage with mature themes, controversial discussions, and provide factual information without hedging or censorship. It maintains strong performance in creative writing, scientific knowledge, and coding tasks, as demonstrated by interactive testing. The model is provided in various GGUF quantizations, making it suitable for local deployment with tools like llama.cpp, LM Studio, or Ollama, even on AMD ROCm-compatible hardware.

Should I use this for my use case?

Good for:

  • Applications requiring uncensored creative writing, including mature themes.
  • Scenarios where direct, unhedged factual responses are preferred, even on controversial topics.
  • Coding assistance and general instruction-following tasks where the base Qwen2.5-7B-Instruct capabilities are desired without refusal behaviors.
  • Local deployment on consumer hardware, especially with GGUF quantizations.

Consider alternatives if:

  • Your application requires strict adherence to ethical guidelines or content moderation.
  • You need a model with explicit safety guardrails against generating potentially harmful content.