The machines
We size the computer around your models, documents and busiest hour. Mac is our usual recommendation for speed and ease of use. AMD Strix Halo and NVIDIA DGX Spark are first-class choices, not custom-build footnotes.
Software and hardware in one quote · exact configuration named in it · path images are renders, per-model photos are the manufacturer’s
The short answer
These are the systems we recommend most often, and for a lot of offices they leave room to spare. If what you actually need is a 16 GB Mac running a small model and transcription, we will quote that instead and say so plainly. The right machine is the one that matches the work — not the largest one we could sell you.
Our usual recommendation
High memory bandwidth, quiet hardware and a clean local-model setup make the Mac the sensible default for most offices.
Mac mini
The current Mac mini range gives us an efficient entry point. M6 configurations suit the compact tier; M5 Pro adds more memory and bandwidth for shared office use.
Mac Studio
Mac Studio is the step up when larger models, longer documents or several active staff make a mini too tight. M5 Max covers many practice workloads; M5 Ultra is for the jobs that need genuinely large unified memory.
Apple memory and internal storage are configuration choices made before purchase. We leave working room for the operating system, document index and model changes instead of quoting a machine that only just fits. Check Apple's current Mac range →
Windows or Linux
A compact route to 64 GB or 128 GB, with a familiar PC operating system and strong value per gigabyte.
Workstation support
HP's small workstation uses the Ryzen AI Max+ PRO 395 and supports Windows 11, WSL2 or Linux. HP states that up to 96 GB of a 128 GB configuration can be assigned to graphics.
High-memory compact PC
The MS-S1 Max is another Ryzen AI Max+ 395 option in 64 GB and 128 GB configurations. It gives us a practical Linux or Windows appliance where the HP support package is not required.
Strix Halo software support is improving, but not every model format or tool is equally mature on Windows, Linux, ROCm and Vulkan. We check the actual workload before recommending it.
NVIDIA and CUDA
A compact 128 GB Linux system for NVIDIA-native work, local AI development and selected larger-model jobs.
DGX Spark uses the GB10 Grace Blackwell superchip and NVIDIA's DGX OS. NVIDIA rates it at up to one petaFLOP of FP4 AI compute with sparsity, and says a single unit can work with models up to 200 billion parameters.
That headline is not the whole buying decision. Its 273 GB/s memory bandwidth is lower than a high-end Mac Studio, so a Spark can hold a large model without necessarily feeling faster when it writes. CUDA compatibility and NVIDIA tooling are the reasons to choose it.
How hardware fits the tiers
Prices are AUD, prices exclude GST, and the written quote is valid for 14 days. See everything included →
Before purchase
Model names and manufacturer numbers are useful, but they do not replace a real workload. We size from your short job, heavy job and awkward job, with memory headroom left for the system around the model.
Manufacturer names
Apple, AMD, HP, Minisforum and NVIDIA product names belong to their makers. We use their public specifications and product photos to explain the hardware we may procure; that does not mean they sponsor or endorse LocallyAI.
Hardware questions
Usually a Mac. Apple silicon gives us a straightforward setup, high memory bandwidth and a well-supported local-model path. We recommend Strix Halo when Windows or Linux is important, and DGX Spark when CUDA or NVIDIA software is the deciding factor.
Yes. Strix Halo systems can be supplied for Windows 11 with WSL2 or for Linux. DGX Spark uses NVIDIA DGX OS. The quoted setup names the operating system and any workload-specific limits.
No. Memory capacity decides what can fit; memory bandwidth, model format and software decide much of how it feels. A slower 128 GB machine can hold a larger model while a smaller Mac answers faster on the model they both fit.
Two DGX Spark systems can be linked for selected NVIDIA-supported workloads. For many offices, separate appliances by department or workload are simpler. We only quote a paired setup after checking the exact model and task.
Storage and memory upgrade options depend on the machine. Apple unified memory is chosen at purchase, while some PC storage can be changed later. Better local models can often be added without replacing the appliance if they fit its tested memory budget.
Not necessarily. We quote the current configuration and availability, then procure against an accepted quote. Public prices are starting points rather than a promise that a particular model is sitting on a shelf.
Yes. Bring us the machine and we will assess what it can genuinely run — memory, storage, operating system, and how much headroom is left once a model is loaded.
If it suits the work, the hardware line comes off the quote entirely and you pay for the software and configuration only: the Harness, the local models, your own templates and the setup. If it is not up to the job, we will say so plainly and tell you what it would take, rather than install something that will disappoint you.
Bring us one job your team repeats every week. We will show you the system doing it →