Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness

Five Qwen3.8-27B models have achieved performance comparable to Fable 5 on the LiveCodeBench hard benchmark, at a fraction of the cost. Opus 5, a previous version of the scaffold, was used in a single pass, with five instances running on a pinned backend. The introduction of ledger-based zero-shot self-orchestration, a training-free method, allowed fresh instances of one model to decompose problems and coordinate through a shared filesystem.

This approach led to gains of up to 25.6 percentage points on pinned backends and offered two routes to frontier-level accuracy. Orchestrated Qwen3.8 Flash Next reached 93.0% pass@1 against Fable 5's 90.4%, while locally served, open-weight Qwen3.8-27B improved from 66.8% to 92.4%. However, not all models benefited equally, with some remaining unchanged or even performing worse.

The gains are attributed to problem decomposition and persistent context, achieved through inference-time organization that can match frontier coding accuracy at a lower cost or slightly exceed it on self-hostable weights. The LiveCodeBench fork, benchmark problem statements, and LaTeX template files each have their own licenses, as detailed in the NOTICE.md file.

The GVS5H report states that five Qwen3.8-27B models achieved Claude Fable 5-level coding performance for a fifth of the price, or on a single GPU.