Start with the constraint, not the brand
Almost every office we quote lands on one of three questions: does it need to be quiet and boring to run, does it need to be Windows, or does it need CUDA. Answer that and the machine chooses itself.
Mac — the default
Mac mini and Mac Studio have high memory bandwidth, run near silent and serve models in the background without anyone thinking about it. Setup is uncomplicated and support is easiest, which matters more than a benchmark when it is sitting in a room with your staff.
This is what we recommend unless something specific rules it out.
Strix Halo — when it has to be Windows
AMD Strix Halo machines reach 128 GB of unified memory in a compact box and run Windows 11, WSL2 or Linux. If your IT is a Windows shop, or your other line-of-business software insists, this is the sensible path and the value per gigabyte is strong.
DGX Spark — when it has to be CUDA
DGX Spark is a 128 GB Linux system built for NVIDIA-native work. Choose it when your team writes code against CUDA, needs the NVIDIA software stack, or works with genuinely large models. Two can be paired for selected workloads.
Memory is the real specification
The single number that decides what you can run is unified memory. Model weights, the document index and the working context all live there at once. We quote headroom rather than the minimum that fits, because a machine that only just fits today is a machine you replace next year.
The software does not change
The Harness, the setup, the support and the price ladder are the same on all three. Compare the three paths in detail, or tell us the work and we will size it.
Book a 20-minute walkthrough and we will run your own document through it.