Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness

Five Qwen3.8-27B models, harnessing GVS5H technology, have achieved results comparable to Fable 5, according to a recent paper. This breakthrough was accomplished using a novel approach called ledger-based zero-shot self-orchestration, which allows fresh instances of a single model to handle tasks and coordinate through a shared filesystem.

This method, when applied to eleven models, resulted in gains of up to 25.6 percentage points on pinned backends and paved the way for frontier-level accuracy. The test model, Orchestrated Qwen3.8 Flash Next, achieved a pass@1 score of 93.0% against Fable 5's 90.4%, while the locally served open-weight Qwen3.8-27B model improved from 66.8% to 92.4%.

The gains were attributed to the decomposition of tasks and the persistence of context during inference. However, not all models saw improvements, as some remained unchanged or even performed worse. The paper emphasizes that inference-time organization can approach frontier coding accuracy at a fraction of the cost, or even surpass it on self-hostable weights.

The code for this method is available under the MIT License, while the paper, figures, and run data are released under CC BY 4.0. The LiveCodeBench benchmark, problem statements, and LaTeX template files have their own distinct licenses.