StudioMemdoor workflow: 5 Qwen3.8-27B models average 9.8 on LiveCodeBench Hard via team collaboration
Original title: Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard
The author implemented the GVS5H multi-agent method as a Memdoor workflow, where 5 Qwen3.8-27B models split into planner, ideator, solver, manager, and repair roles that communicate only through files on disk. On the first 10 LiveCodeBench Hard problems, 5 passes each, it averaged 9.8 passes versus 9.6 for a re-run of the authors' harness and 8.0 for a single Fable 5 call, at a median of 40 minutes and $0.39 per run ($35.79 for all 50). The author notes the sample is small and contamination can't be ruled out, and Fable 5's numbers come from the original authors' files, not this harness. The workflow is 8 YAML files with no model pinned, runs via OpenRouter or locally with Ollama, installable from memdoor.ai/install.sh, code at github.com/guregodevo/memdoor.
The post gives a reproducible multi-agent workflow plus parity data against the paper's harness, so readers can rerun it or adapt it to their own problems.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.