GVS5H: coordinating five model instances to solve hard coding tasks
A shared-filesystem manager harness reportedly lifts open-weight Qwen3.8-27B from 66.8% to 92.4% pass@1 on 100 recent hard LiveCodeBench problems—near the authors’ 90.4% single-call score for Fable 5. The setup lets fresh model instances decompose a problem and coordinate through a persistent plan, notes, and solution; results vary by model, and the headline remains the project’s own benchmark rather than an independent replication.
The one-comment HN thread praised the cost, especially when local hardware is already owned. The repository includes runnable code and results.