GVS5H project claims multi-agent Qwen setup can rival a frontier model on coding benchmark

Coordinating models through a shared ledger

An open-source research project called GVS5H says five instances of the same AI model can improve coding results by dividing work and coordinating through a shared filesystem. The repository describes the method as ledger-based, zero-shot self-orchestration: separate model instances maintain a plan, notes and a current solution without additional model training.

The project's authors evaluated 11 open- and closed-weight models on the 100 latest hard problems in LiveCodeBench. They report gains of as much as 25.6 percentage points for models served on fixed backends, although the benefit was not consistent across every model. Some systems were unchanged or performed worse under orchestration.

For one locally served open-weight configuration, the repository says Qwen3.8-27B rose from 66.8% to 92.4% pass@1. It also reports that an orchestrated Qwen3.8 Flash Next configuration reached 93.0%, compared with 90.4% for a model identified by the project as Fable 5, at 9% of the cost. Those numbers are claims from the project's own benchmark materials and have not been independently reproduced in the supplied evidence.

Inference-time organization, not retraining

The approach is aimed at extracting more capability during inference. Instead of asking one model to solve a programming problem in a single context, fresh instances decompose the task and leave persistent information for one another. The authors attribute the measured improvement to that decomposition and shared context.

This design can trade additional inference calls and coordination overhead for accuracy. Five instances may consume more compute than a single local run even when the underlying weights are relatively compact. Cost comparisons also depend on provider pricing, token usage, hardware utilization and the precise benchmark scaffold. The repository notes that one proprietary comparison used an earlier scaffold and a different serving path, a limitation that complicates direct conclusions.

The code is available under the MIT License, while the paper, figures and run data use CC BY 4.0; benchmark materials retain their original licenses. Running tests requires the project's tooling and access to the chosen model, which may mean a local backend or a provider account.

GVS5H's broader claim is that organization at inference time can narrow the gap between smaller self-hostable models and larger proprietary systems. Its published data offers a reproducible starting point, but wider validation across problem sets, repeated runs and equivalent serving conditions would be needed before treating the headline comparison as a general measure of coding ability.