Xiaomi Launches Local AI Host with Three In-house Chips for On-device Inference

On August 24, Xiaomi unveiled its Local AI Host (an AI Cube prototype), a three-chip system designed to deliver on-device/local AI compute, with large-model inference that works without the cloud.

Confirmed

• The host runs three in-house chips: XRING O3, XRING O100, and XRING D100, supporting local model deployment
• It supports running dual models of 120B and 3B, with a fast/slow system switching mechanism
• The O100 chip uses a 6nm process and Wafer-on-Wafer bonding between NPU and DRAM
• The D100 chip was originally used in Xiaomi's EVs and supports up to 160GB of memory
• The three-chip system delivers 1.2TB/s of bandwidth (one post cited it as roughly the 1.22T level)

Why it matters

• It signals Xiaomi extending its in-house silicon capabilities (smartphone SoCs, automotive chips) into local AI compute hardware, showcasing its on-device large-model strategy
• The O100's wafer-bonded NPU-DRAM design is a fairly advanced packaging approach that could ease the memory bandwidth bottleneck for on-device inference — a highlight noted by multiple commenters
• The fast/slow dual-model architecture (120B+3B) echoes the industry trend of large-small model collaboration for on-device inference

2026-08-24 ~ 2026-08-24 · 7 related posts

• [source] Xiaomi launches local AI host with three custom chips supporting dual models — op7418 · 2026-08-24
• [source] Xiaomi announces AI Cube prototype with 1.2TB/s memory bandwidth — Mysterious_Finish543 · 2026-08-24
• [source] Xiaomi Local AI Host Revealed: Features 6nm Wafer-on-Wafer NPU-DRAM Chip — teortaxesTex · 2026-08-24
• Xiaomi launches new chip supporting local deployment of 200B LLMs — lxfater · 2026-08-24

3 near-duplicate retellings: op7418 · sujingshen · eyishazyer