An open project called Memdoor has published results showing that a coordinated team of five Qwen3.8-27B models outperforms Fable 5 on LiveCodeBench Hard, a demanding coding benchmark. The setup uses a workflow in which multiple model instances collaborate on problems rather than relying on a single large model. The claim has drawn attention because it challenges assumptions about how much performance depends on scale alone.
LiveCodeBench Hard is designed to test code generation on difficult problems, and top scores are typically associated with frontier models. Memdoor's approach appears to emphasize orchestration, shared memory, and division of labor among smaller models. By having several instances propose, critique, and refine solutions, the system can catch errors that a single pass might miss. The repository includes examples of the team-based workflow used for the benchmark.
The result does not mean smaller models are universally better. Benchmark performance depends heavily on prompting strategy, evaluation conditions, and compute budget, and multi-model setups can be more expensive and complex to run. Still, the outcome adds to a growing body of work suggesting that system design can close gaps between model sizes. For developers, it points toward architectures where specialized agents cooperate rather than one model doing everything.
Community reaction has been mixed but engaged, with questions about reproducibility and cost. Supporters see it as evidence that open models can compete when paired with clever engineering, while skeptics want independent verification. Either way, Memdoor has added a notable data point to the debate over whether the future of AI performance lies in bigger models or in better coordination among smaller ones.