A new approach enables complex AI models to run on devices with limited RAM by dynamically loading model weights from storage instead of keeping the entire model in memory. The open source project Turbo Fieldfare implemented this method on Apple Silicon, achieving a sevenfold reduction in memory usage and running a 26-billion-parameter AI model using just 2 GB of active memory. Apple’s “LLM in a Flash” concept allows devices to run models up to twice the size of available RAM. Apple Silicon’s unified memory architecture supports this by enabling efficient data sharing between the CPU and GPU, minimizing latency and maximizing resource utilization. Challenges include reliance on storage speed, thermal management, and hardware configurations, as slower storage or inadequate cooling can impact performance. Dynamic weight loading is most effective with modular AI models and high-memory bandwidth devices, but its applicability is limited for dense, non-modular models, highlighting the need for continued innovation in AI model and hardware design.