I bought the MS-S1 MAX P495 (Ryzen AI Max+ PRO 495, 192GB) to run large AI models locally in my homelab, and after a week of heavy use it has exceeded what I expected.

The 192GB of unified memory is the reason to buy this over the 128GB models. It lets me run 230B to 284B parameter MoE models like MiniMax M2.7 and DeepSeek V4 Flash fully on the GPU, with plenty of room left for long context. My daily driver, Qwen3.8 Flash-Next, generates around 46 tokens per second on code, and gpt-oss 120B does about 55. I use it as a private ChatGPT style server for my family and to run AI coding agents overnight that build and fix my Android apps while I sleep.

Setup on Ubuntu Server 26.04 was smooth. The GPU worked out of the box with the stock Vulkan drivers. You do need to raise the GPU memory limit with a kernel parameter to use most of the 192GB, which is standard for this platform.

Thermals have been great. It lives in a server closet, and under sustained full load it peaked at 78.5C with the GPU clock holding steady and no throttling on the stock fan curve. Dual 10GbE and USB4 v2 are great for a homelab, and the build quality feels premium.

Cons: it is expensive, and there is no fan control exposed in Linux if you want to tune the curve yourself. Neither has been a real problem for me.

If you want to run big AI models locally without a rack of GPUs, this is the box to get.

it's a lovely machine and I've recently updated to the latest BIOS hoping to solve the VRAM allocation limit that fails with AUTO or min/max values for the iGPU. If there was an option to reserve 112GB instead of 96 keeping 16GB of RAM for the CPU this machine could fly on 80/90GB GUFF models via DwarfStar 4 but unfortunately right now only the Spark, among non Apple devices, can run DeepSeek 4 Flash in VRAM, not this box.

I've been using mine for GPU initiated IO development, but also use it for general agentic development and vLLM / lama.cpp testing. Great little machine