Qwen3.5-9B-The-Defiant-Fable
A powerful uncensored LLM
2 min readupdated
Built on Qwen/Qwen3.5-9B by DavidAU and Nightmedia, this new language model packs strong all-around performance into a compact 9.6B parameter footprint. Through multi-stage fine-tuning, it maximizes reasoning ability and instruction following while stripping away content restrictions. It handles complex prompts without refusals and outperforms several much larger models on key benchmarks, making it a solid fit for developers, researchers, and enthusiasts tackling demanding creative or practical tasks.
Key features
- Strong benchmark scores: Matches or beats larger models in the same lineup on core intelligence tests.
- Completely uncensored: Restrictions and content generation refusals are removed (heretic model).
- MTP support: Available in Multi-Token Prediction quantizations for faster text generation.
- Multimodal vision: Retains image input support when paired with an mmproj file.
- Compact reasoning block: Features an optimized internal thinking mechanism for complex logic and coding tasks.
What you can do
- Write, debug, and analyze application source code across different development environments.
- Run advanced roleplay and uncensored creative writing in local clients.
- Tackle complex logical reasoning and analytical problems using the built-in thinking mode.
- Process multilingual prompts, including English and Chinese, with a context window up to 256k tokens.
Example
Recommended generation settings for Thinking mode: temperature=1.0, top_p=0.95, top_k=20, repetition_penalty=1.0.
Read more
DetailsARC-C benchmarks: 0.649 (bf16), 0.640 (mxfp4)
Refusal rate: 6 out of 100 prompts
Context window: up to 256k tokens
Requirements: compatible with standard GGUF clients (LMStudio, KoboldCpp, etc.)










Comments