Xiaomi showed off the Xiaomi AI Cube at today's Xuanjie chip launch event — a prototype AI mini desktop with no word yet on pricing or availability.
It combines three of the company's in-house chips — the Xuanjie O3, O100, and D100 — to pull off a dual-model deployment built around 120B and 3B parameter LLMs, with fast/slow model switching on tap.
The O3 is Xiaomi's flagship phone processor, the O100 is purpose-built to accelerate on-device LLMs, and the D100 handles autonomous driving AI. All three work together in one machine, with the whole system sustaining up to 150W of performance output.
The O100 targets the memory bandwidth bottleneck that plagues large models, delivering 1.22TB/s of near-memory compute bandwidth — 16 times what a typical flagship phone offers. When running Xiaomi MiMo 3B, it can generate up to 330 tokens per second.