zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 21, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated is a 27 billion parameter multimodal language model based on the Qwen3.8 architecture, derived from allura-org/Qwen3.8-27B-Dominatrix. This model has been modified to remove refusal directions from huihui-ai/Huihui-Qwen3.8-27B-abliterated, aiming to retain Dominatrix's roleplay prose while enabling uncensored behavior. It maintains the original vision tower, MTP head, and tokenizer, making it suitable for applications requiring nuanced roleplay and multimodal capabilities without inherent refusal mechanisms.

Loading preview...

Overview

This model, zebulon-prime/Qwen3.8-27B-Dominatrix-abliterated, is a 27 billion parameter multimodal language model built upon the Qwen3.8 architecture. It is a specialized version of allura-org/Qwen3.8-27B-Dominatrix that has been "abliterated" by projecting out the refusal direction found in huihui-ai/Huihui-Qwen3.8-27B-abliterated. The primary goal of this modification is to preserve the original Dominatrix model's roleplay prose while incorporating uncensored behavior.

Key Capabilities

  • Uncensored Roleplay: Designed to maintain high-quality roleplay text generation without the refusal mechanisms present in some models.
  • Multimodal Support: Retains the original vision tower and MTP head from Dominatrix, allowing for image-text processing.
  • Speculative Decoding: The MTP head is usable for MTP speculative decoding, enhancing generation efficiency.
  • Reasoning Preservation: Supports and preserves reasoning capabilities, which can be enabled or disabled via chat_template_kwargs.

Good For

  • Applications requiring nuanced and uncensored roleplay scenarios.
  • Use cases where multimodal input (image-text) is essential.
  • Developers seeking a model with Dominatrix's prose style but without built-in refusal behaviors.
  • Environments leveraging SGLang with DFlash 2 speculative decoding (a specific NVFP4 quant is available).

Usage Notes

This model requires loading with AutoModelForImageTextToText due to its Qwen3_5ForConditionalGeneration architecture; using AutoModelForCausalLM will silently drop vision and MTP tensors. Recommended sampling parameters include a temperature of 1.0–1.25 with min_p 0.1 or top_p 0.95, or alternatively, 0.7 temperature without other parameters. The tokenizer and chat template are unchanged from Dominatrix, including adjusted defaults like preserve_thinking off and reasoning_effort medium.

Finetunes
2 models