BOSGAME M5 AI Mini PC (Ryzen AI Max+ 395)

The BOSGAME M5 puts AMD's Ryzen AI Max+ 395 Strix Halo processor in an all-metal mini desktop. It offers up to 128GB of unified memory and enough I/O to double as a daily-driver desktop. It suits developers and AI hobbyists who want the most local-LLM memory capacity in a compact chassis, without workstation pricing.

Who it's for

• Local inference, RAG pipeline testing and agentic workflow development on models in the 30B–70B range
• Running quantized open-weight models too big for a 16GB or 24GB discrete GPU
• Creators who move footage or datasets off SD cards
• Anyone who wants a quiet desk machine that can switch to full-speed batch jobs

Key features

• AMD Ryzen AI Max+ 395 processor: 16 Zen 5 cores / 32 threads with integrated Radeon 8060S graphics (40 RDNA 3.5 compute units) for gaming and compute-shader work alongside AI acceleration.
• Up to 126 TOPS combined AI performance: The XDNA 2 NPU delivers 50 TOPS on its own; the 126 TOPS figure combines CPU, GPU and NPU throughput.
• 96GB or 128GB LPDDR5X-8000 unified memory: Pick the pool to match your model size; the 128GB configuration loads open-weight LLMs well beyond a typical discrete GPU's VRAM.
• Dual M.2 PCIe Gen4 slots, 2TB included: Fast storage out of the box, with a second slot free as your model library grows.
• Dual USB4 (40Gbps): One front and one rear, for external storage, docking or DisplayPort Alt Mode, plus three USB 3.2 Gen2 Type-A ports and two USB 2.0 ports.
• Quad display support: HDMI 2.1, a full-size DisplayPort with 8K@60Hz, and DisplayPort Alt Mode over the USB4 ports drive up to four displays.
• All-metal chassis with 3-way performance switch: A physical Silent / Balanced / Extreme switch trades noise for sustained throughput, and a bundled vertical stand gives a PS5-style upright footprint.
• Wi-Fi 7 and full-size SD 4.0 card reader: Modern wireless plus a card reader for removable media.

Specifications

Good to know

• Networking is 2.5GbE rather than 10GbE, which matters if you regularly move very large datasets or checkpoints across a local network. Reviewers have flagged this gap at its price and performance tier.
• As with every Strix Halo device, ROCm software support has improved but still trails CUDA for some of the newest inference frameworks, so budget time for setup if you come from an NVIDIA-first workflow.

Choose the M5 when memory capacity per pound and a compact all-metal build matter most. If you need 10GbE or a PCIe slot for a discrete GPU, look at a Ryzen AI Max+ 395 machine that offers them.