lemuralabs/Qwen3.6-27B-Fable-Uncensored
Qwen3.6-27B-Fable-Uncensored by lemuralabs is a 27 billion parameter agentic language model designed for unrestricted, step-by-step reasoning and multi-turn tool-calling. This variant is specifically engineered to have refusal behaviors removed, maintaining top-tier general knowledge with an 88.29% MMLU-Pro score and strong mathematical capabilities at 97.67% GSM8K. It excels in scenarios requiring agentic behavior and complex problem-solving without content restrictions.
Loading preview...
Overview
lemuralabs/Qwen3.6-27B-Fable-Uncensored is a 27 billion parameter agentic model built upon the Qwen3.6 architecture. It is an unrestricted variant of lemuralabs/Qwen3.6-27B-Fable-Agentic, with refusal behaviors significantly reduced or removed. This model is designed for advanced reasoning, agentic tool-calling across multi-turn sessions, and delivering high-quality general knowledge.
Key Capabilities
- Unrestricted Output: Engineered to decline fewer requests, providing more direct responses.
- Top-tier General Knowledge: Achieves 88.29% on MMLU-Pro, demonstrating strong performance in diverse knowledge domains.
- Robust Mathematical Reasoning: Scores 97.67% on GSM8K, indicating proficiency in mathematical problem-solving.
- Agentic Tool-Calling: Supports multi-turn tool-calling, enabling complex interactions and task execution.
- Step-by-Step Reasoning: Utilizes chain-of-thought reasoning to break down and solve problems.
- Multimodal Architecture: Features a multimodal (vision + text) architecture, though specific vision capabilities are not detailed in the provided context.
Benchmarks
Evaluations show that the removal of refusal behaviors did not compromise quality. The model matches or slightly surpasses its base counterpart in MMLU-Pro (88.29%) and maintains strong performance in GSM8K (97.67%) and GPQA-Diamond (56.57%).
Responsible Use
Users are advised that due to reduced refusal behaviors, this model may generate content that a safety-tuned model would typically decline. It is intended for lawful and ethical use, in accordance with its base model and dataset licenses, and should not be used to produce harmful, illegal, or abusive content.