Null-Guard/Qwen3-0.6B-Uncensored

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Null-Guard/Qwen3-0.6B-Uncensored is an 0.8 billion parameter Qwen3-based causal language model developed by Null-Guard. This model is an abliterated version of Qwen/Qwen3-0.6B, with its refusal behavior suppressed via directional ablation of the "refusal direction" in the model's residual stream. It retains the base model's knowledge and reasoning capabilities while offering an unrestricted assistant for research, creative writing, and local/offline use cases.

Loading preview...

Null-Guard/Qwen3-0.6B-Uncensored Overview

This model is an abliterated (activation-ablated) / uncensored version of the Qwen/Qwen3-0.6B base model, developed by Null-Guard. It features approximately 0.6 billion parameters and inherits the Qwen3 architecture and context length. The primary differentiator is the deliberate removal of refusal behavior, achieved through a technique called abliteration.

What is Abliteration?

Abliteration is a method that suppresses refusal behavior in instruction-tuned LLMs without retraining. It involves:

  • Running harmless and harmful prompts through the base model.
  • Identifying a "refusal direction" in the activation space associated with refusals.
  • Ablating (projecting out) this direction from the model's weights at every layer.

This process ensures that the model's general knowledge, reasoning, and capabilities from the Qwen3-0.6B base are preserved, while its tendency to refuse or moralize is significantly reduced. It's a mechanistic intervention, not a values-aligned process.

Key Characteristics & Behavior

  • Uncensored Output: Refusals are substantially reduced, allowing the model to comply with requests that the base model would normally refuse.
  • Preserved Capabilities: The underlying weights and knowledge of the Qwen3-0.6B base model remain unchanged.
  • Small Footprint: As a 0.6B parameter model, it is small, fast, and permissive, though its general reasoning and factual accuracy are limited compared to larger frontier models.
  • Multilingual: Inherits multilingual capabilities from the base Qwen3 model, primarily English.

Intended Use Cases

  • Research: Ideal for studies on alignment, refusal mechanisms, and interpretability.
  • Local/Offline Applications: Suitable for scenarios requiring a small, fast, and unrestricted assistant, such as creative writing, roleplay, or red-teaming systems.
  • Full Control: For users who desire complete control over the assistant's behavior without built-in moralizing or refusals.

Important Note: This model has deliberately reduced safety alignment and may generate offensive, biased, or harmful content. Users are solely responsible for its deployment and outputs.

Finetunes
2 models