AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16
AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16 is a 27 billion parameter Qwen3.6-based model developed by AEON-7, featuring a BF16 precision. This model is a definitively uncensored version of Qwen/Qwen3.6-27B, achieved through a surgical abliteration process that removes alignment overhead while preserving and measurably enhancing core capabilities. It excels at complex reasoning, code generation, and knowledge tasks without refusals, making it suitable for security research, red-teaming, and creative writing.
Loading preview...
AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16 Overview
This model is an uncensored, BF16 precision variant of the Qwen3.6-27B base model, developed by AEON-7. It has undergone a meticulous "abliteration" process to remove safety alignment, resulting in zero refusals on a 100-prompt adversarial battery while maintaining or enhancing core capabilities. The abliteration process, which involved 72 hours of continuous research and tuning with hundreds of AI agents, achieved a KL divergence from the base model under 0.0005, indicating minimal capability damage and even measurable improvements in areas like chain-of-thought commitment and adversarial reasoning.
Key Capabilities
- Uncensored Output: Provides full substantive compliance to prompts, including those typically refused by aligned models (e.g., harmful tools, violence, hate speech).
- Capability Preservation & Enhancement: Statistically indistinguishable from the base model on capability benchmarks, with observed gains in reasoning, code generation, and knowledge tasks due to the removal of the "safety tax."
- High Fidelity: Maintains output length fidelity and internal representation stability, with a KL divergence of 0.000492 from the base model.
- Optimized for Performance: Features a grafted MTP head for speculative decoding (mean accepted length 3.3/3, P0 \u2248 90% acceptance) and is compatible with AEON's vLLM Ultimate container for high-throughput inference.
- Hardware Optimized Variants: While this is the BF16 reference (52 GB), other variants exist for specific hardware, including NVFP4 for Blackwell/DGX Spark and MLX for Apple Silicon, offering different quantization and performance profiles.
Good For
- Security Research & Red-Teaming: Analyzing attack surfaces, vulnerabilities, and failure modes without self-censorship.
- AI Alignment Research: Studying model behavior without imposed guardrails.
- Creative Writing: Generating content without editorial constraints on sensitive topics.
- Jurisdictions with Misaligned Guardrails: Serving users where base model guardrails conflict with local legal frameworks.
- Fine-tuning Workflows: The BF16 precision makes it suitable for further fine-tuning or quant-recipe development.