John Cheung is Assistant Head of IT and Deputy Director of the AI Lab at the College of Professional and Continuing Education (CPCE) of HK PolyU. At CPCE, he contributes to IT leadership, institutional digital transformation, AI governance, e-learning platforms, data services, and the responsible adoption of generative AI in higher education. His professional and research interests include large language models, retrieval-augmented generation, multi-model orchestration, local and cloud AI deployment, and the evaluation of AI-enabled systems. He is the creator and maintainer of OpenFusion, an open-source, OpenAI-compatible framework for experimenting with model-fusion strategies such as voting, parallel synthesis, critique-and-revision, layered refinement, and adaptive workflow planning. His work aims to translate emerging AI research into transparent, reproducible, and practical solutions for education and organizational use.

Large language model performance is usually evaluated at the level of individual models, although practical AI systems increasingly combine multiple models through routing, voting, critique, synthesis, and adaptive orchestration. This paper presents OpenFusion, an open-source, OpenAI-compatible framework for experimentally studying model fusion across local and cloud-hosted language models. OpenFusion supports several bounded inference-time strategies, including parallel synthesis, best-of-N selection, majority and weighted voting, critique-and-revision, layered refinement, fallback routing, and constrained adaptive planning. Using OpenFusion, the study compares fusion strategies with single-model baselines under controlled prompts, model-call budgets, and generation limits. The evaluation considers answer quality, factual accuracy, robustness, response diversity, latency, token consumption, and estimated cost. Particular attention is given to whether independently generated candidate responses provide complementary information, whether synthesis produces a genuinely improved answer rather than merely selecting a fluent candidate, and how performance changes when local and cloud models are combined. The framework also records workflow plans, provider outputs, execution traces, and resource usage to support reproducibility and error analysis. Rather than assuming that additional agents or model calls necessarily improve results, the study examines the conditions under which fusion benefits outweigh increased computational cost and latency. It also distinguishes fixed panel-and-synthesizer architectures from more adaptive multi-model workflows that assign different roles to candidate generators, critics, revisers, and aggregators.The main contribution is a transparent and extensible experimental platform for evaluating model-fusion methods without dependence on proprietary orchestration systems. OpenFusion enables researchers and practitioners to reproduce experiments, introduce new strategies, and compare multi-model systems using consistent interfaces and evaluation procedures.