Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard
Memdoor researchers assembled five instances of the Qwen3.8‑27B language model and evaluated them on the LiveCodeBench Hard benchmark. Their ensemble outperformed the previously leading Fable 5 system, achieving higher scores across the test suite. The result highlights the effectiveness of model ensembling for complex code generation tasks and suggests further potential for scaling similar architectures in future research.