How to choose a mini PC for home AI

A mini PC for home AI is really a memory purchase with a computer attached. Pick the wrong capacity and the box is a frustrating ornament. Pick the wrong kind of memory and you get a machine that loads huge models and then answers at a crawl. This guide gives you a way to decide before looking at a single price.

The two architectures

Unified memory. The processor cores and the graphics hardware draw on one pool. There is no separate VRAM, so a model can use most of the machine's memory. Apple silicon works this way: its MLX framework documentation describes arrays that live in shared memory with no copying between CPU and GPU. AMD's Ryzen AI Max+ and NVIDIA's GB10-based DGX Spark follow the same idea with different engineering.

Discrete GPU. A separate board with its own fast memory, joined to the rest of the PC over a bus. The memory is quick, but the capacity is whatever the card has. The GeForce RTX 5090 carries 32 GB of GDDR7 on a 512-bit interface; NVIDIA's workstation RTX PRO 6000 Blackwell carries 96 GB.

Neither is better in general. The question is which constraint you will hit first.

The unified-memory options on the record

These are manufacturer-published specifications, gathered on 12 November 2026.

Two honest notes. We could not fetch AMD's own product page, which blocks automated access, so we leave out Strix Halo memory bandwidth rather than quote it second-hand. And Apple's Mac mini page lists several models, so check the exact chip-and-memory combination you are buying: the 307GB/s figure belongs to the M5 Pro, not the base M6.

Capacity: choose by the model you want to run

Work backwards from the model. The llama.cpp project's tables show Llama 3.1 70B at 43.1 GB at Q4_K_M, before you add context. So:

• A 32GB machine or card handles models up to the high-20-billion range at 4-bit with modest context. Our worked example puts a 32B model at Q4_K_M with 32k of context at roughly 28 GiB.
• A 64GB machine is where 70B-class models at 4-bit become possible, with room for context.
• 128GB and above is where you stop worrying about the weights and start worrying about speed. Our worked example of the cache shows context length costs memory, too.

This is why the unified-memory boxes matter. A 32 GB flagship graphics card cannot hold a 70B model at all, while the 96 GB workstation card is a different class of product. A 128GB shared-pool machine can, at lower speed.

Bandwidth: the speed side

A bigger pool is no good if it cannot be read quickly, and the published figures vary a lot. The DGX Spark's 273 GB/s and the base Mac mini's 153GB/s sit at the low end for their memory sizes. The Mac Studio M5 Max's 614GB/s and the M5 Ultra's 1.2TB/s are well beyond them. As we show in the CPU, GPU and NPU piece, the theoretical decode ceiling scales directly with this number. A machine with twice the bandwidth can, all else equal, have double the ceiling.

So the framework is:

1. Decide the largest model you will realistically run, and add context.
2. Rule out anything with less memory than that.
3. Among what is left, favour the highest bandwidth you can afford.

Cooling, noise and power

This is where discrete graphics bite. NVIDIA lists the GeForce RTX 5090 at 575 W total graphics power and a 1000 W recommended system power supply. Compare that with Apple's 155W maximum continuous figure for the Mac mini, NVIDIA's 240 W supply and 140 W GB10 for the DGX Spark, or AMD's 45-120W configurable power range for the Ryzen AI Max+ 395. A flagship card is a heating appliance, and fitting one in a small case with quiet fans is a design problem, not a shopping checkbox.

Noise is harder to compare because makers publish it under different conditions. The only primary figure we could verify is Apple's: a typical sound pressure level of 5 dBA at idle for the Mac mini. We could not find a comparable manufacturer-stated figure for the AMD or NVIDIA small-form-factor systems in the sources we fetched, so we make no noise claim about them. Idle is also not load: any machine that runs a model flat out for minutes at a time will be louder than its idle rating. Read reviews that measure the noise under sustained load.

Software support matters as much as hardware

A fast machine with poor software is a slow machine. LM Studio's published requirements tell you where it runs: Apple silicon Macs on macOS 14 or newer, Windows on x64 or Arm, and Linux on x64 or ARM64. AMD's ROCm documentation lists the Ryzen AI Max+ 395 among supported hardware in its 7.2.1 matrix, which matters if you want the AMD GPU stack on Linux. The DGX Spark ships with NVIDIA's AI software stack and, per NVIDIA, can link up to four systems over ConnectX-7 networking, which suits developers more than casual users.

What to buy if...

• ...you want the biggest model for the least fuss, and speed is secondary: a 128GB unified-memory machine. The Spark and the Ryzen AI Max+ class are the two non-Apple routes.
• ...you want the best interactive speed on a model that fits in 32GB: a discrete GPU gives the fastest memory, but plan the case, the supply and the noise.
• ...you live in the Apple ecosystem and want it quiet: a Mac, with memory chosen carefully at the time of order. Apple's spec page lists memory as a configure-to-order option, so decide at the time of purchase and buy for the largest model you expect to run.
• ...you are not sure: buy capacity. You can run a bigger model slowly; you cannot run a model that does not fit.

What we left out

We have not tested any of these machines, so we say nothing about real speeds, fan behaviour or price. Prices move constantly, and we did not verify any in this piece. AMD's Ryzen AI Halo, which AMD said it planned to introduce in the second quarter of 2026 with up to 128GB of unified memory, is not covered in the table because we have not confirmed it is on sale.