Two Years of Local AI on a Laptop: When Open Models Outpaced Moore's Law (huggingface.co)

Between May 2024 and May 2026, significant advancements in local AI capabilities on laptops were demonstrated, with the best-performing open-weight models exhibiting a dramatic increase in intelligence scores while hardware specifications remained largely static. The most advanced model in May 2026, DeepSeek V4 Flash, achieved a score of 47 on the Artificial Analysis Intelligence Index, contrasting sharply with the predicted score of around 20 based on Moore's Law, which states that processing power should double every two years. This marks an astonishing growth rate of 4.7 times in just two years, underscoring the rapid progress in AI model design outpacing traditional hardware improvements. The implications for the AI/ML community are profound, highlighting the importance of software innovations like Sparse Mixture of Experts (MoE) and aggressive quantization techniques that have enabled larger and smarter models to run efficiently on existing hardware. The use of MoE allows massive parameter models to operate with significantly lower active parameters per token, while mixed-precision quantization enhances efficiency without substantial loss in model performance. As the community continues to innovate, future gains will depend on smarter model architectures while navigating the constraints of existing hardware, heralding a potential era of increasingly capable AI running on personal devices.