MS-S1 MAX

Repair shop owner here. I do custom builds as well as sell a lot of mini pcs. if you are looking for a budget AI system. Look at a used workstation with something like an Intel ark pro b550 (16GB). It will run you around $600 depending on the model. Because for a mini pc you are looking around $1000

Any model with Strix Halo 395+ like Minisforum MS-S1-MAX if you have the budget for and can find one. Due to RAM and Storage price increases all models prices jumped through the roof comparing with last year, so you are a little bit late. Remember you need VRAM ( or unified RAM like Strix Halo which

Hi all, about 2 months ago I picked up the MS-S1 MAX and have been having an amazing time adapting my skillset from a web developer to tinkering AI lunatic. I've been custom building an inference platform with Proxmox running an LXC w/ a custom setup that manages "slots" each running a type of strix

DGX is not for anything other than AI, and big models at that. It's a 5070ti speed wise. The 5070ti has 896GB/s memory bandwidth while the DGX Spark and the Minisforum S1 Max have 256GB/s memory bandwidth so it's literally 3.5x slower. You don't buy it for that. You buy it to run 200b parameter mode

My minisforum ms-s1 max 128GB can run 120B models no problem. Generates it faster than you can read it. Using LM Studio. Set it to 96GB VRAM and you still have 32GB system RAM left over. Just did a couple tests. About 45 tokens/sec

Sharing my limited research on workstation. VRAM matters more than speed of the GPU unit -- so newer does not necessary mean better -- as long as the model fits in the VRAm, that will be fastest - tokens per seconds. I believe DGX spark is good option, but workstation PC will be better in general. (

Minisforum MS-S1 Max 128GB (Strix Halo) + RTX 3090 on Oculink eGPU dock. = ~$4500. This configuration is ~100GB of VRAM total (not for a single model), 24GB of it is fast, and it's a solid configuration for fully local orchestration + multiple agent workers. With the new MTP optimizations, 20-25 tp/

Good question, and the answer depends on your hardware. The fault is board-specific, not chip-specific: same silicon runs ROCm fine on a Minisforum MS-S1 Max, dies on my Bosgame M5 (Sixunited AXB35-02 board). Repro on an affected board is trivial: any model load into ROCm triggers it, even a 1.5B. S

Isn’t minisforum ms-s1 max based on the AMD Ryzen AI Max+ 395 combined with a powerful iGPU (AMD Radeon 8060S) - 40 cores ? I guess if you research for LLM inference benchmarks using vulkan on AMD Radeon 8060S - you should be seeing good numbers in terms of output tokens / second - which should be p

For a more detailed NoteBookCheck review One of the most powerful mini PCs of 2025! Minisforum MS-S1 Max review - AMD Strix Halo Power, 128 GB RAM & Radeon 8060S for professionals & AI