How to Run DeepSeek-V3 on a Mini PC: Bosgame M5 AI Cluster Strategy

Running powerful neural networks like DeepSeek-V3 is usually associated with either huge cloud service costs or the purchase of Nvidia H100-level server hardware. However, Bosgame offered a non-standard way, showing that to work with a model of 671 billion parameters, a cluster of seven M5 AI mini-PCs, which is quite scalable.

Instead of server – cluster

The solution is built on combining the resources of seven M5 AI minicomputers running on the AMD Ryzen AI Max+ 395 processor (Strix Halo architecture). The main focus of this chip is on powerful integrated graphics and high memory bandwidth, which is extremely important for LLM operation. Each device in the top configuration is equipped with 128 GB of RAM.

Cluster characteristics

Bosgame’s innovation consists in the use of USB4 Direct Connection technology, which allows you to combine several devices into a connected computing complex without an expensive network architecture.

Technical characteristics of the distributed system:

• Total volume of unified memory: 896 GB.
• Memory available for GPU needs: up to 672 GB.
• No specialized server cooling: devices work in standard office mode.

Price

The pricing of system components has undergone changes. If earlier one M5 AI computer was valued at 1,700 USD (68,475 UAH), at the moment the price has increased to 2,800 USD (112,748 UAH). Experts attribute the price increase to a shortage of AMD’s Strix Halo series chips.

This option is suitable for developers who prioritize data privacy and want to build a local AI infrastructure stack. Despite the increase in price, the use of the M5 AI mini-PC is significantly cheaper than professional server GPU racks, allowing you to scale the power in stages: from a single node to a full-fledged cluster.