The machines we tested: which mini-PC for AI?

The Choosing the hardware guide gives you the method: memory first, bandwidth second, processor last. Here we get to the machines. All of them were reviewed by Frandroid, and nine went through the Frandroid Lab, which measures model speed, power draw and noise. They are sorted by use, from the small server that only hosts the agent up to the AI workstation costing over €6,000.

Tool

Which machine for me?

Three questions, and the reviewed machines that answer them. The rules follow the Choosing the hardware guide: memory decides what fits, budget and system do the sorting.

What do you want to run?

The agent only, models in the cloudSmall models at homeBig models at home

Your budget

Under €1,000€1,000 to €2,000€2,000 to €3,500Over €3,500

System

LinuxMacDoesn't matter

What I suggest

3 machines for you: Apple Mac mini M6, GMKtec EVO-T2S, Minisforum M2.

• 24 GB of unified memory, 17.8 GB of it usable by the graphics processor according to Quelle IA. €1,049, price checked 30 Sept 2026 (starting price (16 GB); the 24 GB reviewed costs more). Frandroid Lab: 33.1 tokens per second on 7B.

  

  Apple Mac mini M6

  Entry level

  Memory

  24 GB unified

  GPU share

  up to 17.8 GB

  Power

  not measured

  Noise

  not measured

  €1,049 checked 30 Sept 2026

  Price: starting price (16 GB); the 24 GB reviewed costs more

  Frandroid Lab 33.1 tokens/s on 7B

  The simplest Mac for local AI: fast and quiet. With 24 GB, 30B models already spill over: aim for more memory to go further.

  • +Quiet even at full load
  • +69 tokens per second on a 3B model in the Lab
  • +Ultra-compact

  • −Only 17.8 of its 24 GB serve the model
  • −Rising price and overpriced options

  Read the Frandroid review In French · 9/10 · 21 Sept 2026 external↗ (opens in a new tab) Frandroid product page In French · specs and prices external↗ (opens in a new tab) Which models can it run? Quelle IA (in French)↗ (opens in a new tab)

• 64 GB of memory. €1,699.99, price checked 29 Jun 2026 (64 GB and 1 TB). Frandroid Lab: 47.1 tokens per second on 3B.

  

  GMKtec EVO-T2S

  Mid-range

  Memory

  64 GB

  GPU share

  not specified

  Power

  70.1 W under load

  Noise

  37 dB

  €1,699.99 checked 29 Jun 2026

  Price: 64 GB and 1 TB

  Frandroid Lab 47.1 tokens/s on 3B

  The most convincing Intel mini-PC for local AI, thanks to its fast memory. For models above 30B, look at the 128 GB Ryzen AI Max+ 395 machines instead.

  • +64 GB of memory at 153 GB/s
  • +Arc B390 integrated GPU, efficient for AI
  • +Quiet and frugal

  • −Soldered memory
  • −Premium price

  Read the Frandroid review In French · 9/10 · 29 Jun 2026 external↗ (opens in a new tab) Frandroid product page In French · specs and prices external↗ (opens in a new tab)

• 32 GB of memory. €1,119, price checked 15 Jun 2026 (32 GB and 1 TB). Frandroid Lab: 43.8 dB.

  

  Minisforum M2

  Entry level

  Memory

  32 GB

  GPU share

  not specified

  Power

  not measured

  Noise

  43.8 dB

  €1,119 checked 15 Jun 2026

  Price: 32 GB and 1 TB

  A well-made desktop mini-PC, but not cut out for local AI with its single memory stick. Add a second stick before running a model.

  • +Efficient Panther Lake processor
  • +Memory upgradable to 128 GB
  • +Compact and quiet

  • −A single memory stick out of the box, which holds memory back
  • −Modest integrated GPU

  Read the Frandroid review In French · 9/10 · 15 Jun 2026 external↗ (opens in a new tab) Frandroid product page In French · specs and prices external↗ (opens in a new tab)

Prices are dated: most go back to the review and may have moved since, sometimes a lot. Check them on the day you buy.

To see which models fit on each machine:Quelle IA, machine by machine (in French)

How these machines are picked

• Agent only: 16 to 32 GB of memory, no graphics card. The machine only hosts.
• Small models: 32 to 64 GB, or a Mac that leaves at least 16 GB to the graphics processor (a 14B fits). On a mini-PC with memory sticks, you need two of them.
• Big models: 64 GB of unified memory or more (Ryzen AI Max, Mac, DGX Spark). An 8 GB graphics card is not enough.
• Budget: the recorded price must stay under the top of the range. A cheaper machine is still offered if it fits.
• Order: first the machines recommended in this guide, then those measured by the Frandroid Lab, then the cheapest. Three at most.

Before you choose: three questions

The agent alone, models in the cloud

This is the cheapest case, and where many of you start. The machine runs the coding agent, your projects, a few containers, and if needed a small model to rephrase or sort things. What matters: silence, low power, a good network. A recycled old PC does the job too.

The NucBox K13 is very thin and almost inaudible (34 dB in the Lab), but its 16 GB are soldered: it will never grow. The NucBox K11 costs a little more and keeps some headroom: 32 GB that can be expanded, an OCuLink port to plug in an external graphics card later, and only 12.6 W at idle. The Lab measured 16.2 tokens per second on a 7-billion-parameter model, enough for small local tasks.

Small models at home (32 to 64 GB)

You want a model at home for everyday tasks, without chasing the giants. An Intel or AMD mini-PC with 32 to 64 GB is enough, on one condition: two memory sticks, or fast soldered memory. A single stick halves the bandwidth, and the integrated graphics processor chokes.

The Minisforum M2 is the desktop mini-PC I recommend to get started, compact and steady under load, but it ships with a single stick: add the second one before you load a model. The GMKtec EVO-T2S goes further with 64 GB soldered at 153 GB/s, the fastest memory the Lab has measured on an Intel mini-PC. The Lab got 47.1 tokens per second on a 3-billion-parameter model, entirely on the integrated graphics processor, for 70 W at most and 37 dB under load.

Big models at home (64 to 128 GB)

This is where local AI gets serious. The chip to remember is the AMD Ryzen AI Max+ 395 (codename Strix Halo): up to 128 GB of unified memory, 96 GB of which the graphics processor can use. It is what runs at my place. Several manufacturers sell it, and the differences come down mostly to cooling, noise and price.

• The GMKtec EVO-X2 is the benchmark: the Lab measured gpt-oss 120B at 35.7 tokens per second, in the “Balanced” BIOS mode. Mind the price on its card: €1,900 was for the 64 GB version with a promo code, at the time of the review; the 128 GB version tested cost more.
• The GMKtec EVO-X3 is just as fast and cools better, but it is bigger and climbed to 186 W under load in the Lab.
• The Framework Desktop costs more, but it is repairable and very well supported on Linux. Its price (€3,889 for 128 GB) was checked on 30 September 2026, without SSD or operating system.

The Minisforum MS-S1 Max uses the same chip. The Lab measured it with very little memory reserved for the graphics processor, hence speeds half as high in the table below: remember to adjust the BIOS. And if you want storage and AI in a single box, the Minisforum N5 MAX is a five-bay NAS with this chip and 64 GB, reviewed by Frandroid (in French). At €2,719 at the time of the review, it is only worth it if you were already shopping for a NAS.

To see which models fit, Quelle IA has one page per configuration (in French): Ryzen AI Max+ 395 with 64 GB (48 GB usable) and with 128 GB (96 GB usable).

The Mac: the simplest route

A Mac mini needs no tuning: macOS lends about three quarters of the memory to the graphics processor on its own. It is quiet, tiny, and Ollama runs very well on it. Its limit is memory: the Mac mini M6 goes up to 32 GB at most.

Frandroid reviewed the 24 GB version, which leaves 17.8 GB to the model according to Quelle IA: enough for a 14B, too tight for a 30B. The Lab measured 33.1 tokens per second on a 7-billion-parameter model. The price on the card (€1,049, checked on 30 September 2026) is for the base 16 GB version; the 24 GB version costs more. For local AI, aim for the M6 with 32 GB (in French), or the M5 Pro up to 64 GB: configurations and prices are in What about a Mac mini?.

Already own a 16 GB Mac mini M4? It makes a very good server for the agent with cloud models, and it runs small models. Do not buy one for local AI: 16 GB is too tight.

The AI workstation: Nvidia’s DGX Spark

The DGX Spark is a developer machine: 128 GB of unified memory, 112 of which the graphics processor can use according to Quelle IA, and above all CUDA, the Nvidia software that nearly the whole AI ecosystem is built on. Several brands sell it under their own name; Frandroid reviewed Dell’s.

The price on the card is that of PNY’s DGX Spark at LDLC, checked on 30 September 2026; the Dell version reviewed was over €7,000. Against a 128 GB Ryzen AI Max+ 395 it is a little faster: on gpt-oss 120B, third-party measurements collected by Quelle IA (in French) give 58.7 tokens per second on the DGX Spark and about 50 on a Framework Desktop, with the same benchmark tool but different versions. About 17% more speed for 57% more money (€6,100 with a 4 TB SSD, against €3,889 without SSD). It makes sense if you develop for CUDA, or want to link several machines together.

What the Lab measurements say

The table lists the machines whose generation speed the Lab measured. In each speed column, the best figure is highlighted.

Frandroid Lab measurements (tokens per second)

Machine	7B	Qwen3 30B (MoE)	32B	70B	gpt-oss 120B	Idle	Load	Noise
GMKtec EVO-X3128 GB	47.8	90.3	11.4	5.3	–	14 W	186 W	48.6 dB
GMKtec EVO-X21128 GB	45.7	92.6	11.4	5.3	35.7	16.6 W	128.9 W	47 dB
Apple Mac mini M624 GB	33.1	–	–	–	–	–	–	–
Minisforum MS-S1 Max2128 GB	22.5	35.2	5	2.4	17.3	–	–	–
GMKtec NucBox K1132 GB	16.2	–	–	–	–	12.6 W	–	34.6 dB
Minisforum MS-0332 GB	8.3	–	–	–	–	23.1 W	–	45.2 dB

What memory buys you

At 32 GB, with modest integrated graphics, the NucBox K11 and the MS-03 were only measured on a 7B, at 16 and 8 tokens per second: a bigger model would fit in memory, but it would crawl. With 128 GB, the EVO-X2 and EVO-X3 load everything the Lab tried, up to the 70B. But look at the speed: a dense 70B drops to 5.3 tokens per second, below the comfort threshold. Memory decides what fits; it does not promise that it will be pleasant.

The real lesson is in two columns. Qwen3 30B is an MoE model, which activates only a small part of itself for each word: it runs at over 90 tokens per second, eight times faster than a dense 32B of similar size. Same for gpt-oss 120B, at 35.7 tokens per second. On a unified-memory machine, these are the models to aim for. More on that in Choosing and sizing your model.

Unified memory or graphics card?

The Lab has not yet measured a mini-PC with a graphics card, so no figures here. The principle is simple: a graphics card is very fast, but it only sees its own memory. An 8 GB RTX leaves 6 GB to the model according to Quelle IA, against 96 GB on a 128 GB Ryzen AI Max+ 395. The card wins on small models; only unified memory loads the big ones.

The Mac mini M6 shows the other side of unified memory: on a 7B it does 33.1 tokens per second, less than the 128 GB Ryzen AI Max machines but twice the K11. It is simply limited by its capacity.

On 24/7: power draw and noise

At idle, the mini-PCs measured draw between 12.6 and 23.1 W. At 25 cents per kilowatt-hour, that is roughly €30 to €50 a year for a machine that waits most of the time. Under sustained load the gap widens: 70 W for the EVO-T2S, 129 W for the EVO-X2, 186 W for the EVO-X3. A machine running at 186 W all year would cost about €400 in electricity, but a home workshop spends most of its time idle.

Noise matters just as much if the machine lives near you. In the Lab, the NucBox K11 (34.6 dB) and the EVO-T2S (37 dB) stay discreet. The Ryzen AI Max machines at full load reach 47 dB (EVO-X2) and 48.6 dB (EVO-X3): you hear them in a quiet room, even if the EVO-X3 stays discreet on short loads.

What the memory crisis changes

The memory shortage has pushed up the price of almost all these machines, and by a lot. A few markers taken from the reviews and from Quelle IA:

• The base Mac mini went from €699 (M4, November 2024) to €1,049 (M6, checked on 30 September 2026).
• Nvidia raised the DGX Spark’s list price from $3,999 to $4,699 in February 2026, citing the shortage.
• The Geekom A9 Max cost €999 when reviewed at the end of 2025; the 2026 edition sold for €1,699 when reviewed in June 2026, with a single memory stick, a choice Geekom justifies partly by the price of memory.

Three words of caution. Treat every price on this page as an order of magnitude, and check it on the day you buy. On a machine with soldered memory (Mac, Ryzen AI Max, DGX Spark), take the capacity you will need right away: you will never be able to add more. On a mini-PC with memory sticks, you can buy 32 GB now and add memory later; nobody knows, though, when prices will come down.