saidutta69/Qwen2.5-3B-Instruct-heretic

Hugging Face
TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 25, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Warm

The saidutta69/Qwen2.5-3B-Instruct-heretic is a 3.09 billion parameter instruction-tuned causal language model, based on the Qwen2.5 architecture, developed by RACER IS OP. This variant is specifically decensored using the Heretic v1.2.0 directional ablation method, which suppresses refusal behavior via targeted weight edits rather than fine-tuning. It is designed for use cases requiring direct answers without refusal, such as local agents, roleplay, or research into alignment mechanics, while largely retaining the base model's knowledge and instruction-following capabilities.

Loading preview...

Overview

The saidutta69/Qwen2.5-3B-Instruct-heretic is a 3.09 billion parameter instruction-tuned model derived from the Qwen/Qwen2.5-3B-Instruct base. Its primary differentiator is the suppression of refusal behavior through a technique called "abliteration" (Heretic v1.2.0). This method involves targeted weight edits to the attention output and MLP down-projections, rather than fine-tuning, ensuring the base model's core knowledge and instruction-following abilities remain largely intact.

Key Capabilities

  • Decensored Responses: Delivers direct answers to prompts that the original Qwen2.5-3B-Instruct might refuse, with only 2 refusals out of 100 adversarial prompts compared to 96 for the base model.
  • Preserved Base Model Integrity: Abliteration minimizes impact on the model's general capabilities, maintaining coherence and knowledge from the Qwen2.5 architecture.
  • Local Runnability: Provided with a comprehensive set of GGUF quantizations, making it suitable for deployment on various consumer GPUs and CPU-only setups.

Good For

  • Local Agents: Ideal for applications where an AI agent needs to provide direct, unfiltered responses.
  • Roleplay: Suitable for scenarios requiring the model to engage in roleplay without generating refusal statements.
  • Alignment Research: Useful for studying refusal mechanics and the effects of alignment techniques by providing a baseline with suppressed refusal.
  • Use Cases Blocked by RLHF: Addresses scenarios where over-refusal from RLHF-tuned models hinders desired functionality.

Finetunes
1 models