AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer0.2K Open Weights Featherless Exclusive Warm

AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is a 27 billion parameter, full-precision BF16 uncensored master model based on Alibaba's Qwen3.8-27B. It has been abliterated for enhanced coherence and direct answers, specifically removing refusal layers while retaining the original vision tower and MTP head. This model is optimized for use cases requiring unfiltered responses, such as security research, red-teaming, and creative writing without content restrictions.

Loading preview...

Model Overview

AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16 is an early access, full-precision BF16 variant of the Qwen3.8-27B model, developed by AEON-7. This 27 billion parameter model has undergone a process called "abliteration" to remove refusal layers and censorship, aiming for greater coherence and directness in its responses. Unlike many public abliterations that prioritize minimizing KL divergence, this model was optimized for "coherence and better answers," accepting a controlled KL drift of 0.0991 nats/token to achieve its uncensored nature.

Key Capabilities & Features

  • Uncensored Responses: Designed to provide direct answers without safety monologues or refusals, even for sensitive queries. It exhibits a "0 outright-refusal" behavior in practice.
  • Qwen3.8 Base: Inherits the capabilities of the Qwen3.8-27B base model, including its vision tower and MTP (Multi-Turn Prediction) head, which remain unmodified.
  • BF16 Full Precision: This is the full-precision master, intended as a base for future quantized versions like NVFP4.
  • Optimized for Coherence: The abliteration process focused on maintaining instructional integrity and useful structure while removing censorship, rather than simply deleting refusal mechanisms.
  • Long Context Support: Validated with a --max-model-len of 16384 tokens, with potential for 262k on suitable hardware.

Intended Use Cases

This model is particularly suited for applications where unfiltered and direct responses are critical, and where the user assumes full responsibility for the content generated. Specific use cases include:

  • Security Research & Red-Teaming: For analyzing vulnerabilities and testing system defenses without model-imposed restrictions.
  • Alignment Work: To understand and explore model behaviors without safety biases.
  • Creative Writing: For generating content without censorship or thematic limitations.
  • Conversations: Engaging in discussions that base models might refuse due to social norms or content policies.

Limitations & User Responsibility

As an uncensored model, it will generate content that the base model would typically refuse, including potentially harmful, illegal, or unethical material. Users are solely responsible for all outputs and must implement their own downstream safety layers and ensure legal compliance. The model does not decide whether to comply; the user does. This is an early access draft, and while functional, it may exhibit minor issues like repetitive loops on extremely long requests, which are being addressed in future iterations.

Finetunes
1 models