junafinity/Qwen-3.8-27B-Uncensored
junafinity/Qwen-3.8-27B-Uncensored is a 27.8 billion parameter multimodal language model, derived from Qwen/Qwen3.8-27B by the Qwen team (Alibaba Cloud). This model has undergone a direct weight edit using ZeroFuse to remove refusal behaviors, making it uncensored without fine-tuning or inference overhead. It is primarily intended as a research instrument for red teaming and defensive cybersecurity, allowing for capability-ceiling measurement and defensive-stack evaluation.
Loading preview...
What is junafinity/Qwen-3.8-27B-Uncensored?
This model is an abliterated version of the Qwen team's (Alibaba Cloud) Qwen3.8-27B, a 27.8 billion parameter multimodal language model. It has been modified using ZeroFuse, a direct weight-editing technique, to remove refusal behaviors without any fine-tuning or inference-time overhead. This process orthogonalizes the refusal direction out of the language model's residual-writing weights, resulting in a standard Hugging Face checkpoint that retains the original model's capabilities, knowledge, and multimodal understanding.
Key Differentiators & Capabilities
- Uncensored Behavior: Refusal behaviors are significantly reduced or eliminated, making it suitable for exploring the model's full capabilities without guardrails.
- Direct Weight Edit: Achieved through a rank-1 projection removal from a subset of the original Qwen weights, not through gradient training or LoRA.
- Multimodal: Inherits the text, image, and video input capabilities of the base Qwen3.8-27B model.
- Research Instrument: Designed specifically for red teaming, defensive cybersecurity research, and interpretability studies.
- Minimal Capability Regression: The abliteration process is optimized to preserve the model's original capabilities, with a low KL divergence from the base model on harmless prompts.
Intended Use Cases
- Red Teaming & Cybersecurity Research: Evaluate the true capability ceiling of the Qwen3.8 weights and test defensive stacks (input filters, output classifiers, moderation APIs) without model-induced refusals.
- Attack Surface Isolation: Facilitate automated red-teaming loops by providing a non-refusing target, isolating the control under test.
- Classifier Training: Generate labeled completions for training or benchmarking output moderation and abuse-detection models.
- Interpretability Research: Study how refusal is represented and where it is written within the model's architecture, as the difference from the base model is a clean experimental control.
Limitations
While significantly uncensored, some refusals may persist (e.g., multi-turn re-assertion, system-prompt-driven refusal, vision-path refusals). It does not add new knowledge or improve accuracy, and users are responsible for implementing their own moderation layers.