From the Mac mini buying rush to the Xiaomi AI Cube and Nvidia DGX Spark, on-device AI is bringing unified memory and local large language model deployment to mainstream users. Yet model weights consume memory, bandwidth constrains generation speed, and questions over model downloads and updates, application invocation, and who maintains continuous operation remain unresolved by a single machine. For now, on-device AI feels more like a "small era," with its ecosystem and usage loop still...

The "small era" of on-device AI began with a Mac mini buying rush. In early 2026, an agent that later became known as "Lobster" drew attention, prompting people to prepare a dedicated computer for an AI that could work at any time. Apple then shifted its messaging toward agents: the new Mac mini began emphasizing round-the-clock agent operation, while Mac Studio continued to sell large memory to users who need to run local large models.

odels are rewriting computer specifications. Two years ago, the typical AI PC configuration was little more than an NPU; Microsoft set the Copilot+ PC threshold at a 40 TOPS NPU, 16GB of memory, and 256GB of storage. But starting with Apple's unified memory, users gradually began deploying local models, and memory became paramount—model weights, context caches, and runtimes quickly devour memory.

This year's products keep growing. The Xiaomi AI Cube prototype packs 80GB of unified memory; Nvidia DGX Spark and a wave of complete systems offer 128GB; AMD's Ryzen AI Max 300 series supports up to 128GB; and the new Mac Studio continues to offer up to 512GB. Using 4-bit quantization as a rough calculation, a