spilol2/Qwen3-0.6B-abliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 14, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

spilol2/Qwen3-0.6B-abliterated is an 0.8 billion parameter causal language model, developed by spilol2, based on the Qwen3-0.6B architecture. This model has undergone an 'abliteration' process to remove refusal behaviors by orthogonalizing refusal directions from its weights, without retraining. It functions as an uncensored instruction-following model, retaining the original Qwen3-0.6B's capabilities but without explicit safety-based refusals. It is primarily intended for research into LLM alignment, red-teaming, and applications requiring custom content policies.

Loading preview...

Overview

spilol2/Qwen3-0.6B-abliterated is a modified version of the Qwen/Qwen3-0.6B causal language model, developed by spilol2. This 0.8 billion parameter model has been processed using an 'abliteration' technique to remove refusal behaviors directly from its weights. Unlike traditional fine-tuning, abliteration modifies existing weights by identifying and orthogonalizing a 'refusal direction' in the activation space, meaning no new data or training is involved. This results in an uncensored instruction-following model that largely retains the original Qwen3-0.6B's capabilities, including its 32768 token context length, but without built-in content refusals.

Key Capabilities

  • Uncensored Output: Designed to generate responses without explicit safety-based refusals, allowing for unrestricted text generation.
  • Retains Base Model Performance: Aims to preserve the reasoning, instruction-following, and knowledge capabilities of the original Qwen3-0.6B model.
  • No Retraining Required: Achieves uncensoring through a geometric modification of weights, making it computationally efficient compared to fine-tuning.
  • Standard LLM Integration: Compatible with common LLM deployment pipelines like vLLM, llama.cpp, LM Studio, and Ollama.

Good for

  • LLM Alignment Research: Ideal for studying refusal mechanisms and interpretability in language models.
  • Red-Teaming & Safety Evaluation: Useful for testing the boundaries and safety of LLM systems.
  • Creative & Roleplay Applications: Suitable for scenarios where the model should not break character or refuse content.
  • Custom Content Policy Enforcement: Developers can use this model to implement their own content moderation at the application layer, rather than relying on the model's inherent safety alignment.

Limitations

While abliteration significantly reduces refusals, it does not guarantee complete removal of all refusal behavior. The model inherits all biases from the original Qwen3-0.6B and its training data. Users are responsible for implementing their own safety guardrails, as this model is not intended for consumer-facing products without proper safety testing.

Finetunes
1 models