DOLGX/QwQ-32B-Preview-abliterated

TEXT GENERATIONPricing:Input $2.72 / Output $4.8Concurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

DOLGX/QwQ-32B-Preview-abliterated is a 32.8 billion parameter language model based on the Qwen/QwQ-32B-Preview architecture. Developed by huihui-ai, this model has been modified using an 'abliteration' technique to remove refusal behaviors, offering an uncensored output. It serves as a proof-of-concept for refusal removal without TransformerLens, primarily for use cases requiring unfiltered language generation.

Loading preview...

Overview

DOLGX/QwQ-32B-Preview-abliterated is a 32.8 billion parameter language model derived from the original Qwen/QwQ-32B-Preview. This version, developed by huihui-ai, has undergone a specific modification process known as 'abliteration' to remove inherent refusal mechanisms present in the base model. The primary goal of this modification is to provide an uncensored language model, demonstrating a proof-of-concept for refusal removal without relying on the TransformerLens library.

Key Capabilities

  • Uncensored Output: Designed to generate responses without refusal behaviors, offering unfiltered content.
  • Proof-of-Concept: Illustrates a method for modifying LLM behavior to remove refusals using an alternative technique.
  • Base Model Performance: Inherits the general language understanding and generation capabilities of the Qwen/QwQ-32B-Preview architecture.

Good for

  • Research into LLM Safety & Alignment: Ideal for researchers studying methods of controlling or altering LLM refusal behaviors.
  • Unfiltered Content Generation: Suitable for applications where uncensored or unrestricted text output is explicitly desired.
  • Exploration of Abliteration Techniques: Useful for developers and researchers interested in the practical application of abliteration for model modification.

Finetunes
1 models