Memdoor Multi-Agent System Beats Fable 5 on LiveCodeBench Hard

A collaborative team of five Qwen3.8-27B models named Memdoor has successfully outperformed Fable 5 on the LiveCodeBench Hard benchmark. This achievement demonstrates the power of multi-agent collaboration using open-source models. This milestone highlights how smaller, open-source models can achieve state-of-the-art performance through structured multi-agent cooperation rather than relying solely on massive standalone systems. It provides a cost-effective and powerful alternative for complex coding and reasoning tasks. The Memdoor framework leverages five distinct instances or roles of the Qwen3.8-27B model to tackle the contamination-controlled LiveCodeBench Hard benchmark. The setup emphasizes how architectural collaboration can overcome the individual limitations of a 27B dense model.