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

saidutta69/Qwen2.5-3B-Instruct-heretic is a 3.1 billion parameter instruction-tuned causal language model based on Qwen2.5-3B-Instruct, developed by saidutta69. This model has been decensored using the Heretic v1.2.0 'abliteration' method, which suppresses refusal behavior via targeted weight edits rather than fine-tuning. It retains the base model's knowledge and instruction-following while directly answering prompts, making it suitable for local agents, roleplay, and research on alignment mechanics where refusal is undesirable. The model supports a 32,768 token context length and is optimized for local deployment on consumer GPUs.

Loading preview...

Overview

saidutta69/Qwen2.5-3B-Instruct-heretic is a 3.1 billion parameter instruction-tuned model derived from Qwen/Qwen2.5-3B-Instruct. Its primary distinction is the removal of refusal behaviors through a technique called abliteration (using Heretic v1.2.0). This method involves targeted weight edits to the attention output and MLP down-projections, preserving the base model's core knowledge and instruction-following capabilities while eliminating its tendency to refuse or lecture.

Key Differentiators

  • Decensored Behavior: Significantly reduces refusals, answering directly even to prompts the base model would decline. Tested with 2 refusals out of 100 adversarial prompts, compared to 96 for the base model.
  • Preserved Core Capabilities: Unlike fine-tuning, abliteration minimally impacts the base model's original knowledge and coherence, as indicated by a low KL divergence of 0.1327 from the base model's output distribution.
  • Local Deployment Optimized: Provided with a full suite of GGUF quantizations (from F16 down to Q2_K), making it highly suitable for running on various consumer GPUs, including those with limited VRAM (e.g., 6GB).

Ideal Use Cases

  • Local Agents: For applications requiring direct answers without refusal.
  • Roleplay: Where models need to maintain character without breaking immersion due to safety guardrails.
  • Alignment Research: For studying refusal mechanics and model behavior without inherent safety filters.
  • Development: For scenarios where RLHF-induced over-refusal hinders desired model interaction.

Finetunes
1 models