saidutta69/Qwen3-0.6B-heretic

Overview

saidutta69/Qwen3-0.6B-heretic is a specialized variant of the Qwen3-0.6B model, created by RACER IS OP using the Heretic v1.2.0 "abliteration" technique. This process involves targeted weight edits to suppress refusal behavior, rather than traditional fine-tuning, ensuring that the base model's core knowledge and instruction-following abilities remain largely intact.

Key Capabilities & Features

• Decensored Output: Significantly reduces refusal rates from 56/100 to 5/100 compared to the original Qwen3-0.6B, allowing it to comply with requests the base model would typically refuse.
• Minimal Divergence: Achieves an exceptionally low KL divergence of 0.0018, indicating that the weight edits are highly precise and do not broadly alter the model's original characteristics.
• Dual-Mode Architecture: Retains Qwen3's unique thinking/non-thinking dual-mode capability, supporting both <think> and direct-answer modes.
• Small Footprint: At 0.6 billion parameters, it is the smallest Qwen3 heretic available, making it efficient for resource-constrained environments.

Ideal Use Cases

This model is particularly well-suited for:

• CPU-only inference and edge deployment: Its small size makes it efficient for running on less powerful hardware.
• Studying refusal mechanisms: Provides a valuable testbed for researchers interested in how refusal behaviors are implemented and suppressed in reasoning-capable models.
• Developers requiring uncensored outputs: For applications where the base model's safety guardrails are undesirable or need to be bypassed for specific research or development purposes.

Important Considerations

Users should be aware that the deliberate suppression of refusal behavior means this model will comply with requests that the base model would refuse, including potentially harmful ones. There is no additional safety filtering. Factual reliability is limited due to its small parameter count, and compliance should not be mistaken for correctness.