5 Qwen3.8-27B agents beat Fable 5 on LiveCodeBench Hard
Original: Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard
Why This Matters
Shows small open models in structured multi-agent loops can rival larger single-model baselines on hard coding benchmarks.
A Memdoor multi-agent workflow using five Qwen3.8-27B models scored a mean 9.8/10 on the first 10 LiveCodeBench Hard problems over five passes, edging the authors' own harness (9.6) and outpacing Fable 5 alone (8.0).
The open-source Memdoor project has published a multi-agent harness for competitive programming problems that beats Fable 5 on LiveCodeBench Hard using only mid-sized open models. Five specialist agents — planner, two ideators, solver, and manager, plus a repair worker — each start with a clean context and communicate exclusively through files on disk. A shell script, not a model, decides whether a solution passes. The workflow reimplements the GVS5H method (Lee, 2026), using the authors' exact prompts and loop on Qwen3.8-27B with a 128K reply cap. Over five passes on the same ten problems used in the original paper, the Memdoor rebuild scored 10/10 on four runs and 9/10 on one, for a mean of 9.8 — versus 9.6 for GVS5H's own harness and 8.0 for Fable 5 running alone. All 50 runs, including every request, reply, grader output, per-call cost, and serving host, are published in bench/results/parity-5pass/. The workflow code was generated by the model itself from a method description on 2026-10-25; no human wrote the files manually.