Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard
Summary
Memdoor documents a multi-agent harness using five Qwen3.8-27B models to tackle programming problems, achieving top performance on LiveCodeBench Hard and outperforming Fable 5. The repository describes five roles (planner, two ideators, solver, manager) and a repair worker, a five-pass evaluation on ten problems, and detailed run logistics across FP8/OpenRouter hosts. It also discusses reproducibility, limitations, and practical guidance to run the workflow.