HauhauCS releases uncensored Qwen3.8-27B-Defiance GGUF
A community contributor has packaged a new version of the Qwen3.8-27B model, combining HauhauCS's uncensored weights with DavidAU's 'Turbo-Brilliance' chat templates. The resulting model, named Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF, is designed to answer user requests without the typical safety refusals, disclaimers, or moralizing found in standard AI outputs.
Key points
- Combines HauhauCS uncensored Qwen3.8-27B weights with DavidAU's 40-mode chat templates.
- IQ4_XS quantization fits on a 24 GB GPU with 131K context window.
- Removes safety refusals and disclaimers, allowing direct answers to sensitive prompts.
The release includes 40 distinct 'construct' modes, such as Joker, Dr. House, and a Socratic tutor, which users can activate via specific tags in their prompts. The contributor verified that all 866 tensors match HauhauCS's original uncensored files byte-for-byte, ensuring the base weights are unchanged. The template modifications specifically remove instructions for 'final safety checks' and 'ethical guardrails,' replacing them with directives to answer fully and directly.
Technical testing on an RTX 4090 using the IQ4_XS quantization showed decode speeds of 86-90 tokens per second in instruct mode. The model requires approximately 24 GB of VRAM for the IQ4_XS version with 131K context, while the Q6_K_P version needs around 26 GB for weights alone. The project is released under the Apache-2.0 license, and the contributor notes that while the model is text-only, users are responsible for complying with local laws regarding generated content.
Qwen3.8-27B-Defiance-Uncensored-MTP-GGUF
huggingface.co · 17 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- NVIDIA releases Nemotron 3.5 Super VL and Bolt edge models for enterprise agents · 10 src
- Meta launches Meta Spark Open, optimized for Apple Mac Studio · 1 src
- Three Jinja chat templates for Qwen3.8-Flash-Next · 1 src
- OpenAI releases GPT-Live-1.1 voice model with full-duplex support · 1 src
- Llama.cpp merges support for IBM Granite Switch 4.1 · 1 src
Comments
via GitHub Discussions