Five Qwen Models Using File-Based Teams Beat Major Benchmarks

Coordinated through Memdoor YAML workflows, a team of five Qwen3.8-27B models collaborating via disk files successfully tackled LiveCodeBench Hard.

A groundbreaking multi-agent experiment has demonstrated that smaller open-weight models, when organized correctly, can rival massive monolithic systems in complex programming tasks. Utilizing the Memdoor workflow framework, a team of five Qwen3.8-27B models—comprising a planner, two ideators, a solver, a manager, and a repair worker—collaborated exclusively by reading and writing files on disk.

Operating with fresh contexts and governed strictly by an external evaluation script rather than model logic, the team solved challenging problems on LiveCodeBench Hard. Each complete 5-pass benchmark run required a median time of 40 minutes at a compute cost of just $0.39, proving the immense viability of modular, file-swapping agent architectures over costly proprietary APIs.