@jiqizhixin: "A 120B model. On your desk. At 150 watts." — Xiaomi's answer to the local-AI question is a box with three of its own c…

X AI KOLs Timeline Products
xiaomi ai-hardware chips local-inference desktop-ai

Summary

Xiaomi launched its XRING chip lineup and a desktop AI Cube prototype capable of running large language models locally, with specialized chips for AI acceleration and autonomous driving.

"A 120B model. On your desk. At 150 watts." — Xiaomi's answer to the local-AI question is a box with three of its own chips inside. Xiaomi didn't launch a chip last night. It launched three. Aug 24, Beijing, the new XRING lineup: - XRING O3 — 2nd-gen flagship SoC, TSMC 3nm, 24B transistors (up from 19B). Single-core near A19 Pro, multi-core catching M4, GPU close to M5. - XRING O100 — 6nm on-device AI accelerator. 14-core NPU, 1.22TB/s memory bandwidth (~16× a mainstream phone), 330 tokens/s local inference. - XRING D100 — 3nm autonomous-driving-class chip. 20-core CPU + 16-core NPU, up to 160GB on a single die, enough to run a 200B+ model locally. Then the "one more thing": Xiaomi AI Cube, a prototype desktop box packing all three. 150W, runs a 120B LLM at 30 tokens/s. O100 and D100 ship commercially next year.
Original Article
View Cached Full Text

Cached at: 08/25/26, 06:11 PM

“A 120B model. On your desk. At 150 watts.” — Xiaomi’s answer to the local-AI question is a box with three of its own chips inside.

Xiaomi didn’t launch a chip last night. It launched three.

Aug 24, Beijing, the new XRING lineup:

  • XRING O3 — 2nd-gen flagship SoC, TSMC 3nm, 24B transistors (up from 19B). Single-core near A19 Pro, multi-core catching M4, GPU close to M5.
  • XRING O100 — 6nm on-device AI accelerator. 14-core NPU, 1.22TB/s memory bandwidth (~16× a mainstream phone), 330 tokens/s local inference.
  • XRING D100 — 3nm autonomous-driving-class chip. 20-core CPU + 16-core NPU, up to 160GB on a single die, enough to run a 200B+ model locally.

Then the “one more thing”: Xiaomi AI Cube, a prototype desktop box packing all three. 150W, runs a 120B LLM at 30 tokens/s.

O100 and D100 ship commercially next year.

Similar Articles