Zhaoxin deploys DeepSeek-R1 models across entire CPU lineup, from desktops to servers

Zhaoxin has just shared that it’s successfully rolled out the DeepSeek-R1 Distill models across its entire range—from everyday desktop PCs to high-powered servers.

The Chinese CPU maker announced that its processors now support several versions of the DeepSeek-R1 Distill model, whether it’s the 1.5B, 7B, 14B, 32B, 70B, or even the colossal 671B variant. This rollout spans all of their x86-compatible hardware, showcasing the flexibility of their technology.

Plus, Zhaoxin’s CPUs natively support Linux, Windows, and various homegrown Chinese operating systems while also being fully compatible with local GPU AI acceleration cards. This creates a robust computing platform ready for all sorts of AI tasks.

Desktop AI Deployment

The KaiXian KX-7000 desktop processor demonstrated a successful local deployment of the DeepSeek-R1-7B model when paired with domestically produced graphics and acceleration cards. By tapping into DeepSeek’s advanced language understanding and reasoning, this setup enables features like smart document handling, data analysis, and coding assistance, all smoothly integrated with applications such as WPS Office and VSCode.

The KX-7000 is part of Zhaoxin’s latest generation of desktop chips. It features the “Century Avenue” microarchitecture with innovative chiplet interconnection, packs 8 CPU cores with 32MB of shared L3 cache, and can hit speeds of up to 3.7GHz.

Workstation Performance

The Lianhe Donghai XRS302 workstation, powered by the KH-40000/16 server processor, successfully deployed the DeepSeek-R1-7B, 14B, and 32B models. Outfitted with four high-performance domestic GPU inference acceleration cards, this workstation delivers cost-effective yet robust AI inference capabilities.

Zhaoxin has fine-tuned its CPU-GPU interconnect system to increase data throughput while cutting latency. The workstation also benefits from a low-power, quiet design—perfect for office settings that need on-premises AI inference without noise and energy overhead.

Server-Grade AI Processing

On the higher end, the Kaisheng KH-40000/32 processor features a dual-socket setup with 64 total cores and support for up to 2TB of memory. Teamed up with domestic accelerator cards, this CPU comfortably deployed the DeepSeek-R1-70B model.

The KH-40000/32 also managed to deploy the full DeepSeek-R1-671B model entirely on its own—without external acceleration cards. This feat underscores just how capable Zhaoxin’s server CPUs are in effortlessly handling models with hundreds of billions of parameters.

While Zhaoxin's claims are impressive, these performance results should be taken with a grain of salt. As with any vendor-provided benchmarks, real-world performance may differ from controlled test environments, and the company may present optimized scenarios that highlight their products in the best possible light.