Beelink SER9 PRO AMD Pro Ryzen™ AI 9 HX 370 (OpenClaw & Local LLM Pre-installed)

The Beelink SER9 PRO pairs AMD’s Ryzen AI 9 HX 370 (12 Zen 5 cores, a Radeon 890M iGPU, and an XDNA 2 NPU good for 80 AI TOPS) with OpenClaw and local LLM tooling pre-installed. It sits in the same compact 135 x 135 x 44.7mm chassis as Beelink’s other SER-series machines. It is for developers and AI enthusiasts who want on-device model inference in a machine small enough to disappear on a desk, not a full tower workstation.

Who it's for

• Developers building agentic coding workflows with OpenClaw who want no API latency, no per-token cost, and no code or prompts sent to a third party
• Running mid-size quantized open-weight models locally with Ollama or llama.cpp
• Multi-monitor productivity setups with up to triple 4K displays
• General-purpose use as well, including creative work and light gaming on the 890M iGPU

Key features

• AMD Ryzen™ AI 9 HX 370, 12 cores, 24 threads: Zen 5 cores boost up to 5.1GHz with 36MB of combined L2+L3 cache, for compiling, running local models, and heavy multitasking at once.
• AMD Radeon™ 890M, RDNA 3.5, 16 compute units: Running at up to 2900MHz, it is widely regarded as the fastest integrated GPU class in a mini PC, able to handle GPU-accelerated inference alongside light gaming.
• 80 AI TOPS via XDNA 2 NPU: A dedicated NPU gives OpenClaw and local inference tooling like Ollama or llama.cpp on-device throughput for quantized models instead of depending on a cloud connection.
• 32GB LPDDR5X 7500MHz memory: High-bandwidth soldered memory keeps model loading and inference responsive, sized for mid-size quantized open-weight models alongside your normal workload.
• Dual PCIe 4.0 NVMe storage, up to 8TB: Two M.2 2280 PCIe 4.0 x4 slots support up to 8TB combined, and the retail unit includes a 1TB or 2TB Crucial SSD, leaving room for multiple model checkpoints.
• Triple 4K display output: DisplayPort 1.4 (4K@240Hz), HDMI 2.1 (4K@240Hz), and USB4 (40Gbps/TB3/PD/DP1.4) together support up to triple 4K displays.
• 32dB near-silent cooling: Efficient bottom-intake airflow keeps the chip stable at 65W TDP with very low fan noise, even during sustained AI inference.
• AI-enhanced voice pickup and dual speakers: A built-in microphone processed through an AI chip gives 360° barrier-free voice recognition within 5 meters, with dual speakers for calls and media.

Specifications

Good to know

• The 80 TOPS NPU makes local inference practical for smaller quantized models run entirely on-device.
• For the largest local models, 70B-class and above, the GTR9 Pro’s 128GB unified memory platform is the better fit.

Choose the SER9 PRO for mid-size quantized models and agentic coding workloads in a quiet, compact box. Step up to the GTR9 Pro only if you need the largest models.