The project "smolmaxxing" aims to maximize the ability to run large language models (LLMs) on small and old devices, such as TV boxes, old phones, and single-board computers. By utilizing custom kernels, smarter memory use, and weight streaming, the project has successfully run models with up to 16 billion parameters on devices like a 3rd-gen Xiaomi TV Box and a 13-year-old iPad Air, achieving impressive performance metrics, including 24-26 tokens per second on the iPad Air with a 1.7B-parameter model, demonstrating the potential to unlock the hidden capabilities of everyday devices.