164 points57 comments
Five Qwen3.8 models orchestrated to beat Claude on coding benchmark
What is Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness?
GVS5H lets five instances of the open-weight Qwen3.8-27B model split a coding problem, share notes via a common folder, and solve it together. The approach hits 93% pass@1 on LiveCodeBench Hard, beating Claude’s 90.4% while costing only about 9% of the proprietary model.
Who it is for
Developers who need near-state-of-the-art code generation without paying for closed-source APIs can run it locally.