Memory decides which model you can use
A language model only answers quickly if it fits entirely in the memory the graphics uses. In a unified memory mini PC, the processor and the integrated graphics share the same memory, so the graphics can use most of it. That is why a small computer runs models that would need a very expensive graphics card in a tower.
The sum is simple: a model compressed to 4 bits takes about 0.6 GB per billion parameters. A 30-billion one takes about 19 GB and a 70-billion one about 42 GB. On top of that comes the memory of the conversation, which grows with the length of the documents.
The three sizes and what fits in each
The memory is soldered to the board, so it cannot be upgraded later: you have to choose well when buying. These are the three computers and their price today.
| Computer | Memory for graphics | Models that fit | What for | Price |
|---|---|---|---|---|
| Compact AI 32 | Up to 24 GB | Up to about 19 GB | Assistant for a small team, transcription, light vision | €2,065 |
| Compact AI 64 | Up to 48 GB | Up to about 42 GB | Large models, assistant on documents, agent development | €2,940 |
| Compact AI 128 | Up to 96 GB | Up to about 80 GB, or several at once | The largest local models, several models, long analyses | €4,820 |
What speed to expect
Unified memory is fast, but not as fast as that of a dedicated graphics card. A mid-sized model answers fluently. A 70-billion-parameter one writes a few words per second: fine for one person waiting for the answer, not for twenty at once.
If many people will use it at the same time, or if the answer has to be immediate, a dedicated graphics card performs better. That is why these computers accept one.
When to add a professional graphics card
Any of the three can be fitted with an NVIDIA RTX PRO and it stays compact: it moves to an enclosure of about 10 litres. It pays off when there are many simultaneous users, several cameras running machine vision or software that requires CUDA, the NVIDIA technology used by almost every training tool.
| Graphics card | Memory | Added to the computer |
|---|---|---|
| NVIDIA RTX PRO 2000 Blackwell | 16 GB with ECC | + €2,005 |
| NVIDIA RTX PRO 4000 Blackwell | 24 GB with ECC | + €3,820 |
| NVIDIA RTX PRO 4500 Blackwell | 32 GB with ECC | + €5,760 |
The model can sit entirely on the graphics card, to gain speed, or be shared between the card and the unified memory, to gain size.
Which software is used
Open models run with free programs such as Ollama, LM Studio or llama.cpp, and are queried from the browser through an interface similar to that of cloud assistants. They work on Windows and on Linux. On Linux the integrated graphics can use a little more memory than on Windows.
The computer is delivered with Linux installed free of charge, or with Windows 11 Pro on request. If you wish, we leave it with the models downloaded and the interface ready to open.
What to check before buying
Besides memory, four things separate a working machine from a hobbyist one: 2 TB of storage or more, because each model takes tens of gigabytes; remote management over IP, to look after it without being in front of it; a warranty that covers continuous use; and that someone has tested it for 24 hours at full load before shipping.
Related hardware and guides
Frequently asked questions
Are 32 GB enough for an AI assistant?
For a small team, yes: models up to about 19 GB fit, and they answer questions about documents and write fluently. If you want larger models or several at once, the 64 GB one is the balance point.
Can I upgrade the memory later?
No. In these computers the memory is soldered to the board, which is what makes it so fast. Storage can be expanded through a second M.2 slot, and graphics with a dedicated NVIDIA RTX PRO.
Unified memory mini PC or a tower with an RTX?
The mini PC wins when the model is large and few people use it: it loads what does not fit on a 16 or 24 GB card. The tower with an RTX wins on speed and with many simultaneous users, and is needed to train models.
How much power does it draw?
A fraction of what a workstation with a dedicated graphics card uses: it has a low-power processor and a 450 W power supply working well below its limit. That is why it also suits installations with limited power.
What will you use it for?
Tell us which models you want to run, how many people will use it and where it will work. We will propose the compact, with or without a professional graphics card, and its price within 12 working hours.