Uncensored Power: Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive

By vramkickedin | | 2 min read
A jagged stone made of dark matte obsidian with deep sharp cracks of the stone and a piercing, brilliant white light emits softly.

HauhauCS has released an uncensored variant of the Qwen3.6 (shortly after their 3.5 variant Qwen3.5-9B-Uncensored-HauhauCS-Aggressive) a language model that removes standard content restrictions. The system processes inputs through an expert-routing architecture, activating just a fraction of its thirty-five billion parameters per task to maintain efficient performance.

Independent developers often struggle to find reliable local models that generate complete answers without blocking requests. This release delivers an open-weight set that preserves original performance benchmarks while allowing unrestricted text generation.

Model Size: from 11.7GB & VRAM GPU: requirements vary

Optimized local processing and custom precision settings

  • Removes built-in filters to produce complete responses across all prompt categories.
  • Routes only eight specialist experts per token to lower memory demands.
  • Offers custom K_P precision files that improve output quality with minimal size increases.
  • Handles image and video input when paired with the included projection file.

Professionals managing private datasets can run this software on standard desktop computers. Keeping computation offline maintains full data control while supporting drafting and analysis workflows.

Developer notes on quantization and system behavior

The creator removed safety filters for the aggressive edition, though occasional short disclaimers may still appear due to original training data. Custom quantization files rely on an importance matrix to protect critical weights. Users might see placeholder symbols in certain interfaces, which do not affect performance.

Emphasizing the goal behind the project,

"These are meant to be the best lossless uncensored models out there,"

noted the developer in a post. Teams running extended reasoning tasks must enable specific template flags and allocate a large context window to avoid output truncation.

Access the complete model archive directly from the project hosting page.