AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is a 27 billion parameter language model developed by AEON-7, based on Alibaba's Qwen3.8-27B. This full-precision BF16 master has been "abliterated" to remove refusal layers, enhancing coherence and directness in responses while retaining its vision tower and native MTP head. It is designed for use cases requiring uncensored outputs, such as security research, red-teaming, and creative writing, and is validated on NVIDIA H200 with vLLM.
Loading preview...
Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 Overview
This model is a 27 billion parameter, full-precision BF16 master developed by AEON-7, derived from Alibaba's Qwen3.8-27B. It has undergone a specialized "abliteration" process to remove inherent refusal behaviors and censorship, aiming for enhanced coherence and directness in its responses. Unlike many other uncensored models, this version prioritizes maintaining the model's instructional integrity and useful structure rather than simply minimizing KL divergence, which can often lead to degraded performance.
Key Capabilities
- Uncensored Responses: Provides direct answers to queries that base models typically refuse, making it suitable for sensitive or unrestricted content generation.
- Coherence-Optimized: Abliteration focused on improving answer quality and directness, rather than just removing safety layers, resulting in a more coherent output.
- Retains Base Model Strengths: The vision tower and native MTP (Multi-Turn Prediction) head from the original Qwen3.8-27B remain untouched, preserving its multimodal and conversational capabilities.
- Speculative Decoding Support: Custom-trained AEON DFlash2 drafter is available for this model family, offering significant speed improvements (e.g., +21% faster than stock DFlash2 at T=0.6).
- Deployment Flexibility: Available in various optimized versions, including NVFP4-MIXED for DGX Spark and RTX GPUs, and MXFP4-MXFP6 for AMD ROCm hardware.
Good For
- Security Research & Red-Teaming: Ideal for exploring model vulnerabilities and generating content without artificial restrictions.
- Creative Writing & Roleplay: Enables unrestricted narrative generation, free from content filters or moralizing responses.
- Alignment Research: Useful for studying model behavior when safety alignments are removed and for developing new alignment techniques.
- Complex Problem Solving: Designed to provide direct answers to hard questions without evasiveness or refusal, even on topics like tools, chemistry, or exploit-shaped code.
- High-Performance Inference: Validated on NVIDIA H200 with vLLM, supporting features like reasoning parsers and auto-tool choice for efficient deployment.