mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16 is a 27 billion parameter language model based on the Qwen3.8 architecture, specifically re-ablated from the official Qwen/Qwen3.8-27B base. This model is engineered to minimize refusal and deflection behaviors, achieving 92.0% direct compliance on harmful prompts in 'thinking-OFF' mode. It retains the original Qwen3.8's multi-token prediction capabilities and a 32768 token context length, making it suitable for applications requiring direct and unfiltered responses.

Loading preview...

Model Overview

The mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16 is a 27 billion parameter model derived from the Qwen3.8 base, specifically designed to address and correct 'deflection' and 'refusal' behaviors found in its predecessor. This version, 2.1.0, was developed using a direct-compliance objective, optimizing for direct answers rather than keyword-based refusal or steering requests towards 'safer alternatives'.

Key Capabilities & Features

  • Uncensored Responses: Achieves 92.0% direct compliance on harmful prompts when run in 'thinking-OFF' mode, significantly reducing deflection and refusal rates compared to its base model and prior iterations.
  • Robust Ablation Method: The model underwent a re-ablation process focusing on attn.o_proj and mlp.down_proj layers, minimizing multi-token KL divergence and non-direct response counts.
  • Knowledge Retention: Evaluation benchmarks show strong retention of knowledge and reasoning capabilities, with MMLU, ARC-Challenge, HellaSwag, and Winogrande scores within ~1.3 points of the base model.
  • Multi-Token Prediction (MTP): Retains the Qwen3.8's multi-token prediction speculative-decoding head, enhancing generation efficiency.
  • Context Length: Supports a 32768 token context length.

Important Usage Notes

  • 'Thinking-OFF' Mode: For optimal uncensored behavior and accurate performance on tasks like GSM8K, it is crucial to run the model with enable_thinking=False.
  • Vision Caveat: The ablation was performed on text residuals; multimodal decensoring is not claimed or validated.

Quantizations

Various GGUF quantizations are available, ranging from IQ2_M (9.8 GB) to Q8_0 (28 GB), all retaining MTP capabilities.

Finetunes
1 models