Xiaomi Unveils Three-Chip AI Cube Prototype Running 120B Models Locally

On August 24, Xiaomi unveiled a Local AI Host (AI Cube prototype) with a three-chip system for cloud-free on-device LLM inference. Multiple posters (@op7418, @eyishazyer, @sujingshen, @markk) confirmed the core configuration; @julianharris views it as a potential Chinese-market competitor to Nvidia's GB10. It remains a prototype with no launch details.

Confirmed

• The host runs three in-house chips: Xring O3, Xring O100 and Xring D100, enabling local model deployment
• Supports parallel 120B + 3B dual models with fast/slow system switching; demos ran fully locally
• 200 TOPS NPU compute, 1.22TB/s AI memory bandwidth, 150W sustained performance (@markk)
• O100 uses 6nm with Wafer-on-Wafer NPU-DRAM bonding, noted by @teortaxesTex as the standout technology
• D100 originally shipped in Xiaomi's EV; system supports up to 160GB memory
• @lxfater relays Xiaomi's claim of local 200B-parameter model support and speculates about in-car local model deployment
• Xring O3, Xiaomi's second flagship SoC, is 3nm with Antutu score over 5.22M; AI integrated across CPU (new matrix engine), GPU (first neural acceleration unit) and NPU (@量子位)

Unconfirmed

• The 200B local deployment claim appears in a single post and differs from the widely reported 120B dual-model setup; attribution of the 200 TOPS figure (whole system vs. O3's NPU) varies across posts

Why it matters

• It signals Xiaomi extending its chip portfolio (phone SoC, automotive silicon) into local AI compute hardware, forming an edge-AI matrix from phones to desktop hosts
• The Wafer-on-Wafer NPU-DRAM bonding is an advanced packaging route that could ease memory-bandwidth bottlenecks in on-device inference
• The 120B+3B fast/slow architecture echoes the industry trend of large-small model collaboration, seen by @eyishazyer as a step toward practical edge AI; if the GB10 comparison holds, competition in local AI compute will intensify
• The 150W / 160GB combo indicates a desktop-class device; real-world performance awaits productization

2026-08-24 ~ 2026-08-25 · 11 related posts

Full story(2 episodes)→

• Episode 1: Xiaomi Unveils Three-Chip AI Cube Prototype Running 120B Models Locally(2026-08-24, 11 posts)
• Episode 2: Xiaomi Unveils Xring O100 Edge AI Chip, Commercial in 2027(2026-08-24, 3 posts)

• [source] Xiaomi launches local AI host with three custom chips supporting dual models — op7418 · 2026-08-24
• Xiaomi announces AI Cube prototype with 1.2TB/s memory bandwidth — Mysterious_Finish543 · 2026-08-24
• 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
• Xiaomi Unveils Xuanjie O3 Flagship SoC with All-Module AI — 量子位 · 2026-08-24
• [source] Xiaomi AI Cube Prototype: 200 TOPS NPU, 160GB Unified Memory, Runs 120B Models Locally — julianharris · 2026-08-24

5 near-duplicate retellings: op7418 · sujingshen · eyishazyer · mark_k · ccerrato147