Minisforum MS-S1 Max

AMD's Strix Halo answer to the unified-memory argument Apple has been winning — and it lands in a chassis small enough to put in a bag. 128GB of soldered LPDDR5x addressable by the accelerator means models that would otherwise need a multi-GPU rig run on one small machine drawing a fraction of the power.

The architectural trade, again

The catch is the one that always applies here: memory bandwidth and compute sit well below a discrete GPU, so you are trading tokens per second for the ability to load the model at all.

• On a 70B-class model, that trade is obviously worth it — nothing else this size and this quiet will run it.
• On a 13B model that already fits in a 5070 Ti's 16GB, it is obviously not — the discrete card generates far faster.

The trade-offs

• Soldered memory. The configuration you buy is the configuration you keep — no upgrade path.
• Speed. Acceptable for interactive use on large models; not for high-throughput serving.

Who should buy

Edge deployments that need a large model in a small box, local open-model work where capacity beats speed, and buyers who want serious capability without a tower. If your target models already fit in consumer VRAM, buy the GPU instead.