AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16
AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16 is a 27 billion parameter, BF16 precision, uncensored variant of Alibaba's Qwen3.6 model, featuring a 32K context length. It has undergone a specialized abliteration process to remove safety alignments while preserving or enhancing core capabilities, achieving zero refusals on adversarial prompts with minimal KL divergence from the base model. This model is designed for use cases requiring full substantive compliance, such as security research, red-teaming, and creative writing without editorial constraints, offering enhanced reasoning and calibration honesty.
Loading preview...
Overview
AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16 is a 27 billion parameter, BF16 precision model derived from Alibaba's Qwen3.6, featuring a 32K context length. This model has been meticulously engineered through a process called "abliteration" to remove safety alignments and censorship, resulting in a model that provides full substantive compliance to user prompts. The abliteration process, which involved 72 hours of continuous research and tuning with hundreds of parallel AI research agents, ensures that core capabilities are not merely preserved but measurably enhanced, with a KL divergence from the base model under 0.0005.
Key Capabilities
- Uncensored Responses: Achieves 0 refusals on a 100-prompt adversarial battery, providing full compliance even for prompts typically refused by aligned models.
- Capability Preservation & Enhancement: Maintains core reasoning, math, code, and knowledge capabilities, with KL divergence from the base model at 0.000492, indicating no significant capability damage. Some metrics, like NatInt reasoning, have shown improvements in similar abliterated models.
- Enhanced Reasoning: Exhibits longer, more committed chains of thought, improved adversarial-example reasoning, and cleaner calibration on contested topics due to the removal of the "safety tax."
- Multimodal Support: The base architecture supports multimodal inputs, and specific variants preserve the vision tower.
- Speculative Decoding: Ships with the original
mtp.*head restored from the base model, enabling Multi-Token-Prediction (MTP) speculative decoding with high acceptance rates (e.g., mean accepted length 3.3/3, P0 ≈ 90% acceptance).
Good For
- Security Research & Red-Teaming: Ideal for analyzing attack surfaces, vulnerabilities, and failure modes without self-censorship.
- Alignment Research: Useful for studying AI alignment and understanding model behavior without imposed guardrails.
- Creative Writing & Roleplay: Provides an environment for creative expression without editorial constraints.
- Jurisdictional Compliance: Suitable for serving users in regions where standard model guardrails may conflict with legitimate local legal frameworks.
- Hardware-Specific Deployments: This BF16 variant is recommended for A100/H100 (80 GB) or RTX PRO 6000 (96 GB) for fine-tuning or as a full-precision reference. Other optimized variants (NVFP4, MLX) are available for specific hardware targets like DGX Spark, Blackwell, or Apple Silicon.