⬢github Python · 1.1K ★ +299 since we first saw it · pushed 21 h ago · MIT
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
GVS5H is a training-free orchestration harness that runs multiple fresh instances of a single LLM, coordinating them through a shared filesystem (plan, notes, solution) to boost coding benchmark accuracy. On hard LiveCodeBench problems, orchestrated Qwen3.8-27B models reportedly match or exceed Claude Fable 5 at a fraction of the cost, or fully open-weight on one GPU.
Why now: It's trending on Hacker News and gaining stars because of claims that five cheap open-weight models, managed by this harness, match a frontier proprietary model on LiveCodeBench Hard — a striking cost/performance result backed by a published paper.
Who it is for: ML researchers and engineers benchmarking LLM coding ability, and teams seeking frontier-level code generation using open-weight models on limited budgets or local hardware.
Stars over our 242 snapshots: 768 to 1.1K, since 2 d ago.
Where people talked about it
API: https://socialmediatrends-api.osmike.com/v1/repos/slee-persis/GVS5H