bzannah/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
bzannah/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is a 27 billion parameter Qwen3.8-based language model, fine-tuned by AEON-7 for enhanced coherence and directness in responses. This BF16 model is specifically abliterated to remove safety alignment drag, resulting in an uncensored model that provides more direct answers across a wide range of topics. It retains the original Qwen3.8 vision tower and MTP head, making it suitable for security research, red-teaming, and creative writing without content restrictions.
Loading preview...
Model Overview
bzannah/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is a 27 billion parameter model derived from Qwen/Qwen3.8-27B, meticulously abliterated by AEON-7. The primary goal of this modification was to enhance response coherence and directness by removing the 'safety alignment drag' inherent in many base models, rather than simply minimizing KL divergence. This results in a model that provides more straightforward answers, even on topics typically refused by aligned models.
Key Characteristics
- Uncensored Nature: The model is designed to provide direct answers without internal refusals, making it suitable for sensitive or restricted topics. It explicitly states that it will write content the base model would refuse, including tools, chemistry, exploit-shaped code, violence, sexuality, and ideologies.
- Coherence-Optimized Abliteration: Unlike models that prioritize zero KL drift, this model was optimized for "coherence and better answers," accepting a controlled KL drift (0.0991 nats/token) for improved output quality.
- Zero Outright Refusal: Benchmarking shows the model has 0 outright refusals on harmful, sexual, and harmless held-out prompt sets, with any 'judge-R' classifications often being preambles or disclaimers rather than true refusals.
- Full-Precision BF16: This release is the full-precision BF16 reference, with an NVFP4 sibling planned for future release.
- Retained Capabilities: The original Qwen3.8 vision tower and MTP (Multi-Turn Prediction) head remain unmodified and fully functional.
- Validated Performance: Verified on NVIDIA H200 with vLLM 0.27.1, supporting text, vision, and speculative decoding with MTP.
Intended Use Cases
- Security Research & Red-Teaming: Ideal for exploring model vulnerabilities and conducting red-team exercises without artificial content restrictions.
- Alignment Work: Useful for studying and developing new alignment techniques by understanding the behavior of an unaligned model.
- Creative Writing: Enables creative writing without a 'hall monitor,' allowing for exploration of a broader range of themes and narratives.
- Conversations on Refused Topics: Facilitates discussions and content generation on subjects that base models typically refuse due to social norms or publisher policies.
Users are explicitly responsible for all outputs and downstream actions, and are expected to implement their own safety layers.