zeeksa/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-BF16

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

zeeksa/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-BF16 is a 27 billion parameter Qwen3.6 model, dequantized from HauhauCS's uncensored GGUF finetune into a BF16 safetensors format with a 32768 token context length. This model includes a rebuilt vision tower for full image/video input support and an MTP head for speculative decoding. It is optimized for uncensored applications requiring multimodal capabilities within the PyTorch/Hugging Face ecosystem.

Loading preview...

Model Overview

This model, zeeksa/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-BF16, is a 27 billion parameter Qwen3.6 variant, specifically a BF16 safetensors version of the HauhauCS uncensored finetune. It was created by dequantizing the highest-fidelity public GGUF artifact (Q8_K_P) back into the official Qwen/Qwen3.6-27B Hugging Face layout. While the original finetune was GGUF-only, this version provides a PyTorch-compatible format, maintaining a quality ceiling near Q8.

Key Capabilities & Features

  • Uncensored Finetune: Based on the HauhauCS aggressive uncensored finetune, suitable for applications requiring less restrictive content generation.
  • Multimodal Support: Includes a rebuilt vision tower, enabling full image and video input capabilities. The mmproj components were re-fused and Conv3D patch-embed re-stacked.
  • Speculative Decoding: Features an MTP (Multi-Turn Prediction) head, taken from the stock Qwen/Qwen3.6-27B, to facilitate speculative decoding in compatible engines.
  • Standard Qwen3.6 Architecture: Utilizes the tokenizer and configurations from the stock Qwen/Qwen3.6-27B, ensuring compatibility with the base model's architecture, including its gated attention mechanism.

Usage Considerations

  • Format: Provided in BF16 safetensors, ideal for the Hugging Face/PyTorch ecosystem, including transformers, vLLM, and further finetuning or quantization.
  • Resource Requirements: Requires approximately 54 GB of VRAM, making it suitable for multi-GPU setups or environments where single-GPU serving is handled by FP8 or GGUF variants.
  • Engine Compatibility: Due to Qwen3.6's new gated attention, users should ensure they are using recent versions of transformers (>=5.10) or vLLM for full architectural support.

Finetunes
1 models