---

Qwen3.6-27B-Fable-Uncensored is an unrestricted 27B agentic model that reasons step-by-step and operates tools across multi-turn sessions, while delivering top-tier general knowledge. It is a refusal-reduced (abliterated) variant of lemuralabs/Qwen3.6-27B-Fable-Agentic.

Benchmarks

Evaluated head-to-head against the base model on identical questions, with an 8192-token budget so step-by-step reasoning is never truncated.

Uncensored at no cost to quality. Removing refusals left capability fully intact: it matches the aligned model on MMLU-Pro (88.29%), GSM8K, and hard GPQA-Diamond reasoning (56.57% vs base 57.58%, within noise) - and still beats the base on knowledge.

MMLU-Pro by category

Category	Base	This model
Biology	96.0%	98.0%
Business	78.0%	86.0%
Chemistry	84.0%	78.0%
Computer Science	84.0%	82.0%
Health	82.0%	88.0%
Math	94.0%	96.0%
Physics	90.0%	90.0%
Overall	86.86%	88.29%

Highlights

• Unrestricted - refusal behavior removed; declines far fewer requests
• Top-tier knowledge - 88.29% MMLU-Pro, above the base, with zero abliteration tax
• Strong math - 97.67% GSM8K
• Agentic tool-calling + step-by-step reasoning

Responsible use

This model has had refusal behavior reduced, so it may produce content a safety-tuned model would decline. You are responsible for how you use it - use it lawfully and ethically, in accordance with the base-model and dataset licenses. It is not intended for producing harmful, illegal, or abusive content.

Capabilities

• Unrestricted, agentic multi-turn tool-calling
• Chain-of-thought reasoning
• Top-tier general knowledge and math
• Multimodal (vision + text) architecture

Usage (vLLM)

vllm serve lemuralabs/Qwen3.6-27B-Fable-Uncensored --trust-remote-code --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --max-model-len 16384

Tip: this model reasons before answering - give it a generous max_tokens so it can finish its chain of thought.

License

Apache 2.0