HFCK99/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
HFCK99/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is an uncensored, 27 billion parameter variant of Alibaba's Qwen3.8-27B model, optimized for direct answers without safety-induced refusals. It features a 32768 token context length and retains the original vision tower and MTP head. This model is specifically designed for use cases requiring unfiltered responses, such as security research, red-teaming, and creative writing, by removing internal refusal layers while preserving coherence.
Loading preview...
Model Overview
HFCK99/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is an early access draft of an uncensored version of the Qwen3.8-27B model. Developed by AEON-7, this 27 billion parameter model aims to provide direct answers to a wide range of queries by removing the base model's refusal layers, without sacrificing coherence. It maintains the original Qwen3.8-27B's vision tower and Multi-Turn Prediction (MTP) head, offering a 32768 token context window.
Key Differentiators
- Uncensored Responses: The primary goal is to eliminate safety-induced refusals, allowing the model to answer questions that the base model would typically decline or lecture about. This is achieved through a surgical "abliteration" process rather than a simple deletion of refusal layers.
- Coherence Over Zero KL: Unlike many public abliterations that prioritize minimizing KL divergence, this model optimizes for coherence and better answers. A slight KL drift (0.0991 nats/token) is considered acceptable as it reflects the model's shift to more direct responses.
- Zero Outright Refusals: In practical testing, the model exhibits zero outright refusals, even on harmful or sensitive prompts. Judge-R flags are often due to preambles or disclaimers rather than actual refusal to provide the requested content.
- Retained Capabilities: The model retains the core capabilities of the Qwen3.8-27B base, including its vision processing and MTP functionalities, which were untouched during the abliteration process.
Intended Use Cases
This model is designed for scenarios where unfiltered and direct responses are critical:
- Security Research and Red-Teaming: For analyzing vulnerabilities, generating exploit-shaped code, or understanding system behaviors without artificial constraints.
- Alignment Work: For studying model behavior when safety alignments are removed.
- Creative Writing: For generating content without censorship or moralizing interventions.
- Conversations Beyond Social Norms: For exploring topics that base models typically refuse due to publisher-defined social norms.
Users are explicitly responsible for implementing their own safety layers and ensuring legal and ethical compliance, as the model itself does not perform internal refusal.