Xiaomi AI Cube and Xring O100: 1.22 TB/s, 330 Tokens/s and 120B Local AI

Xiaomi has shown the AI Cube Prototype, a compact local-AI system built around three Xring processors: O3, O100 and D100.

The system sustains up to 150 W and runs a 120B + 3B dual-model configuration locally. Its dedicated AI accelerator, Xring O100 (玄戒 O100), combines 6 nm logic with vertically stacked DRAM and delivers 1.22 TB/s of near-memory bandwidth.

O100 reaches up to 330 tokens/s on Xiaomi MiMo 3B.

Xiaomi AI Cube specifications

O3 provides the general-purpose CPU/GPU/NPU platform, O100 handles high-bandwidth AI inference, and D100 adds Xiaomi's large-memory high-compute architecture.

Xring O100 specifications

3D-stacked logic and DRAM

O100 places DRAM directly above the AI logic to create a short, wide path between model weights and the NPU.

The stack combines:

1. one 6 nm logic die;
2. two DRAM dies;
3. Wafer-on-Wafer vertical assembly;
4. Hybrid Bonding;
5. Face-to-Face metal connections;
6. 28,672 effective data lines.

The bonding pitch is 1.4 μm and the TSV diameter is about 0.7 μm. Xiaomi's packaging comparison increases the data-path count from 96 lines in a conventional PoP-style reference to 28,672 lines in O100.

1.22 TB/s near-memory bandwidth

O100 provides 1.22 TB/s between its stacked memory and AI compute subsystem.

Large-model inference repeatedly reads model weights during token generation, making memory bandwidth a major part of inference throughput.

O100 reaches its bandwidth through the vertical memory stack and dense interconnect rather than a conventional external memory bus.

330 tokens/s on Xiaomi MiMo 3B

Xring O100 runs Xiaomi MiMo 3B at up to 330 tokens per second.

A 3-billion-parameter model occupies approximately:

Xiaomi also demonstrated an O3 + O100 dual-chip prototype, with O3 handling system-level work and O100 handling large-model inference.

Xuanwu high-bandwidth matrix bus

O100 uses Xiaomi's Xuanwu high-bandwidth matrix bus to connect the NPU cores and memory subsystem.

It supports parallel processing across the 14 NPU cores, dynamic routing and high-throughput data movement between compute and the stacked DRAM.

120B + 3B dual-model AI Cube

The AI Cube runs two model sizes locally:

• 120B-class large model
• 3B-class small model

Xiaomi describes the system as a fast/slow dual-model design. The smaller model handles fast-response workloads while the 120B model provides the larger parameter budget.

A 120B model at 4-bit uses about 60 GB for weights. Inference also uses memory for KV cache, runtime buffers and activations.

AI Cube compared with other local-AI systems

The closest product-class comparisons are compact high-memory systems such as NVIDIA DGX Spark and AMD Ryzen AI Halo. Mac Studio and discrete GPU systems provide useful bandwidth and memory reference points.

Four RTX 5090s

A four-card RTX 5090 workstation provides:

• 128 GB aggregate VRAM
• 7.168 TB/s aggregate local VRAM bandwidth
• 2,300 W total GPU board power

Each card has its own 32 GB VRAM pool, with multi-GPU model partitioning handled across PCIe.

Four Radeon AI PRO R9700s

A four-card R9700 workstation provides:

• 128 GB aggregate GDDR6 VRAM
• 2.56 TB/s aggregate local memory bandwidth
• 764 TFLOPS aggregate peak FP16 matrix throughput
• 1,200 W total GPU board power

R9700 is a general-purpose workstation GPU for ROCm and professional GPU workloads, while O100 is a dedicated high-bandwidth inference accelerator.

2027 rollout

Xring O100 is planned for commercial use in 2027. Xring D100 is also scheduled for 2027, while O3 reaches the Xiaomi 18 Fold and Pad 9 Pro Max in September 2026.

The AI Cube is the first public system combining all three.

Read next: Xiaomi Xring O3: full specifications, LPDDR6 and benchmarks and Xiaomi Xring D100: 3nm, 160GB unified memory and 200B models.