Self-Orchestrated Models Match Frontier Performance at Fraction of Cost

GVS5H introduces a training-free coordination method where multiple instances of a single model work together to solve coding problems. Each instance decomposes problems and shares progress through a shared filesystem, allowing persistent context and collaborative refinement without requiring model retraining.

Across eleven models tested on the hundred hardest LiveCodeBench problems, the method delivered improvements up to 25.6 percentage points. Orchestrated Qwen3.8 Flash Next achieved 93.0% accuracy versus Fable 5’s 90.4% while costing only 9% as much. An open-weight Qwen3.8-27B model improved from 66.8% to 92.4% when locally deployed. Transcript analysis showed decomposition and persistent context drove the gains.

Performance improvements were inconsistent across all tested models, with some showing no benefit or worse results. The approach offers two paths to frontier-level accuracy: using orchestrated commercial models economically, or self-hosting open weights on modest hardware like a single GPU. Code is released under MIT license.