slee-persis/GVS5H
GVS5H: Five Qwen3.8-27B Models Match Claude Fable 5 on LiveCodeBench Hard | Fable 5 Level Coding for a Fifth the Price - or on a Single GPU
What it solves
Frontier coding performance usually requires large proprietary models that are expensive. This project offers a training-free alternative: by coordinating several instances of the same model, it can match or even exceed the accuracy of a top-tier closed model on hard coding benchmarks, at a fraction of the cost.
How it works
The method, called ledger-based zero-shot self-orchestration, runs fresh instances of one model to decompose each problem and coordinate through a shared filesystem. That filesystem holds a plan, notes, and the current solution, so the instances work together without any fine-tuning or extra training. It was tested on the 100 most recent hard LiveCodeBench problems across eleven open and closed-weight models. Running the code requires uv and an API key for the desired model; you can choose single-call or manager-based (--engine multiagent) execution.
Who it's for
Researchers and practitioners who want to improve coding model accuracy without retraining, especially those interested in using self-hostable, open-weight models to get frontier-level results at low cost.
Highlights
- Gains of up to 25.6 percentage points on pinned backends across the tested models.
- Orchestrated Qwen3.8 Flash Next reaches 93.0% pass@1 vs. Claude Fable 5's 90.4%, at 9% of the cost.
- Locally served Qwen3.8-27B improves from 66.8% to 92.4% pass@1.
- Gains are not universal: some models show no improvement or even degrade.
- The code is MIT-licensed; the paper and run data are under CC BY 4.0.
Related
- Dispatch
- Project
- Project
- Dispatch
- Dispatch