Memdoor's five-Qwen3.8-27B team beats Fable 5 at LiveCodeBench

Memdoor's team-solve example runs five Qwen3.8-27B instances as a planner, two parallel ideators, a solver and a manager, with a repair worker on top, and scored a mean 9.8 of 10 over five passes on the first ten LiveCodeBench Hard problems. That beats the 8.0 that Fable 5 posts on the same ten alone, and edges past the 9.6 of the GVS5H harness by Lee (2026) that it reproduces.

The agents never share a chat context.

"Five roles (planner, two ideators, solver, manager) and a repair worker each start from a fresh context and talk only through files on disk," the README states. The planner writes a plan file, the solver writes solution.py, the manager reads the test logs, the repair worker patches, and each agent re-enters cold with only the files it needs. The model is Qwen3.8-27B with a 128K reply cap, and the workflow reuses the GVS5H authors' exact prompts, so the comparison reads as an implementation test, not a prompt-engineering win.

Fifty runs cost $35.79, or $38.77 counting 16 incomplete attempts, and consumed 15.3 million output tokens. A median run came in at forty minutes and $0.39.

The authors flag contamination up front: the contests "date from 29 March to 6 April 2025, and the model may have seen them in training: contamination cannot be ruled out." Memdoor lands in a run of open-weights Qwen experiments on our radar, alongside Lyra's Evil-Qwen trick and the Nesformer NES-in-Qwen project, both filed within the last two days.