saidutta69/Qwen3-0.6B-heretic

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwen3-0.6B-heretic is a 0.8 billion parameter decensored variant of Qwen/Qwen3-0.6B, created by saidutta69 using the Heretic v1.2.0 directional ablation method. This model suppresses refusal behavior through targeted weight edits, preserving the base model's knowledge and instruction-following capabilities. It is designed for developers seeking Qwen3's dual-mode architecture without refusal guardrails, making it suitable for CPU-only inference, edge deployment, and studying refusal mechanisms.

Loading preview...

Qwen3-0.6B-heretic: A Decensored Qwen3 Variant

Qwen3-0.6B-heretic is a specialized version of the Qwen/Qwen3-0.6B model, developed by saidutta69. This model has undergone a process called "abliteration" using the Heretic v1.2.0 tool, which involves targeted weight edits to suppress refusal behavior. Unlike traditional fine-tuning, this method aims to keep the base model's core knowledge and instruction-following abilities largely intact while removing its safety guardrails.

Key Characteristics & Performance

  • Decensored Behavior: Refusal rates were significantly reduced from 56/100 in the original model to 5/100, indicating a high degree of compliance with requests the base model would typically refuse.
  • Minimal Knowledge Degradation: The model exhibits an exceptionally low KL divergence of 0.0018 compared to the original, suggesting that the targeted edits had a very narrow impact on the model's overall distribution and capabilities.
  • Dual-Mode Architecture: Retains Qwen3's unique thinking/non-thinking dual-mode capability, allowing for flexible response generation.
  • Resource Efficient: At approximately 0.6 billion parameters, it is the smallest Qwen3 heretic variant available, making it suitable for environments with limited computational resources.

Ideal Use Cases

  • CPU-Only Inference: Its small size makes it practical for deployment on CPUs or edge devices.
  • Refusal Mechanism Research: Provides a testbed for researchers interested in studying how refusal mechanisms operate in reasoning-capable language models.
  • Unfiltered Content Generation: For developers who require a model that will comply with a broader range of prompts without built-in refusal guardrails. Users are responsible for its deployment and output.

Finetunes
1 models