Zero's new mini PCs pack 56 GB of total RAM, split between the host and a discrete NPU card
Chinese mini PC maker Zero (零刻) has listed two new "AI PC" desktops on JD.com that split computing between two physically separate subsystems: a standard x86 host that handles the OS and daily applications, and a discrete NPU accelerator card with its own dedicated 24 GB LPDDR5 memory pool for local model inference. The NPU card is rated at 160 TOPS and comes preloaded with an "OpenClaw" local inference stack running on Ubuntu.
Specifications
The two SKUs differ only in their host CPU and I/O. The flagship SEi14 AI pairs a Core Ultra 9 185H (16C/22T, up to 5.1 GHz, 24 MB L3) with integrated Arc graphics (8 Xe cores, 2.35 GHz) and Thunderbolt 4. The cheaper SEi13 AI drops to a last-generation Core i7-13620H (10C/16T, up to 4.9 GHz, 24 MB L3) with UHD Graphics (64 EUs, 1.5 GHz) and loses Thunderbolt in favor of standard 10 Gbps USB-C. Both configurations carry the same NPU card and the same 24 GB of dedicated inference memory, and support the same W4A8/W8A8 quantized inference formats.
| Model | SEi14 AI | SEi13 AI |
|---|---|---|
| Host CPU | Intel Core Ultra 9 185H, 16C/22T, up to 5.1 GHz, 24 MB L3 | Intel Core i7-13620H, 10C/16T, up to 4.9 GHz, 24 MB L3 |
| Integrated GPU | Intel Arc, 8 Xe cores, 2.35 GHz | Intel UHD Graphics, 64 EU, 1.5 GHz |
| System memory | 32 GB DDR5 SO-DIMM (2x slots, up to 5600 MT/s) | 32 GB (16 GB x2) LPDDR5 |
| Storage | 1 TB M.2 2280 PCIe 4.0 x4 (up to 4 TB) | 1 TB M.2 2280 PCIe 4.0 x4 (up to 4 TB) |
| NPU card | Discrete, 160 TOPS | Discrete, 160 TOPS |
| NPU dedicated memory | 24 GB LPDDR5 | 24 GB LPDDR5 |
| NPU features | Local model weights, W4A8/W8A8 quantized inference, text and multimodal processing, long-context calculation | Same |
| Networking | WiFi 6 (AX200), Bluetooth 5.2 | Same |
| I/O | USB2.0 x2, USB3.2 10Gbps x2, 2.5G LAN, 3.5mm audio x2, USB-C 10Gbps (data), DP1.4a, HDMI, Thunderbolt 4 (40Gbps/PD/DP1.4) | USB2.0 x2, USB3.2 10Gbps x2, 2.5G LAN, 3.5mm audio x2, USB-C 10Gbps (data), DP1.4a, HDMI, USB-C 10Gbps (data and video) |
| Dimensions | 135 x 135 x 44.7 mm | 135 x 135 x 44.7 mm |
| Adapter | 19V/6.32A | 19V/6.32A |
| Price (32GB/1TB/24GB NPU config) | CNY 13,988 (~USD 2,067) | CNY 11,895 (~USD 1,757) |
The architecture is notable for keeping inference entirely off the host CPU and iGPU. Task submission and the return of results between the host and NPU card are handled over an internal link, with the NPU card handling its own model weights, quantization pipeline, and multimodal processing independently of whatever the host OS is doing. This mirrors the growing trend of NPU-first local inference boxes but is unusual in that it dedicates a full 24 GB memory pool exclusively to the accelerator rather than sharing system RAM.
Both models are listed on JD.com and carry a five-year warranty with a 30-day trial period, though both configurations are currently listed as out of stock.
Add as a preferred source on Google
