denyzhirkov/Qwen3.8-27B-AEON-ULTIMATE-U-BF16
denyzhirkov/Qwen3.8-27B-AEON-ULTIMATE-U-BF16 is an uncensored, BF16 variant of Alibaba's Qwen3.8-27B model, featuring 27 billion parameters and a 32768-token context length. This model has undergone a specialized "abliteration" process to remove refusal layers, aiming for enhanced coherence and direct answers without safety monologues. It is optimized for use cases requiring unfiltered responses, such as security research, red-teaming, and creative writing, while retaining the base model's instructional capabilities.
Loading preview...
Model Overview
denyzhirkov/Qwen3.8-27B-AEON-ULTIMATE-U-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, with a 32K context window, has been specifically modified to remove inherent refusal behaviors and censorship, aiming for more direct and coherent responses. The vision tower and native MTP head remain unmodified from the base Qwen3.8-27B.
Key Differentiators
- Uncensored Output: The primary goal of this model is to provide answers without safety monologues or refusals, even for sensitive queries. It is designed to be a "0 outright-refusal" model.
- Coherence-Optimized Abliteration: Unlike methods that blindly minimize KL divergence, this model's abliteration process (using
abliterix 1.12.2) prioritized coherence and answer quality over simply matching the base model's token distribution. This results in a KL drift of ~0.10 nats/token, which is considered a sign of genuine unlock rather than damage. - Retained Base Capabilities: The modification aims to remove censorship without "lobotomizing" the model, preserving the instructional structure and quality of the original Qwen3.8-27B.
- BF16 Precision: This release is in BF16 (bfloat16) full precision, serving as the master for potential future quantized versions.
Intended Use Cases
This model is designed for scenarios where unfiltered responses are critical, including:
- Security Research and Red-Teaming: For generating content that aligned models would typically refuse.
- Alignment Work: To study model behavior without safety constraints.
- Creative Writing: For unrestricted content generation without internal censorship.
- Conversations: For topics that the base model might avoid due to social norms.
Users are explicitly warned about the unaligned nature and are solely responsible for implementing downstream safety layers and ensuring legal and ethical compliance.