What computer do you need to run an AI on your own data files?
On a Windows PC you need a graphics card with about 12 GB of its own memory. On a Mac you need an M4 Pro chip or better (an M3 Max or M3 Ultra also qualifies), with 18 GB of memory or more. Most computers sold today, including recent and expensive ones, do not have that. The rest of this page explains why, what smaller machines can run, and how to check yours in one minute.
Before you buy anything, test. The numbers on this page say whether the model can run at all, not whether it will feel fast enough to you. That depends on the machine, on your files and on how you work, and what one person finds comfortable another finds too slow. The free trial runs the real model on your own hardware, with no time limit. Use it before you pay for a license, and, if you are thinking of a new computer for this, try Totsum on a machine like it before you order one. During the private beta, the trial is closed and the beta takes its place. The download page explains what the beta is, and you can apply there.
Why memory matters more than the processor
A language model is one big file of numbers. To answer a question, the whole file has to sit in memory that the chip can read very fast, and the chip reads through all of it for every word it writes. Two things follow.
The file has to fit. A model with 12 billion parameters, compressed the usual way, is a file of about 7 GB, and it needs a few more gigabytes of room to work on your question and your table. That is where the 12 GB figure comes from.
Speed depends on how fast the memory is, more than on how fast the processor is. On most PCs, the fast memory is the one on the graphics card, called VRAM. The ordinary RAM of the computer is much slower for this job. So a laptop with 16 GB of RAM and no separate graphics card does not qualify, however good it is at everything else. On a Mac with an Apple Silicon chip, one pool of memory is shared by the processor and the graphics part, and it is fast enough. That is why a Mac counts its whole memory and most PCs count only the graphics card’s. A new kind of PC works like a Mac, with one shared pool of fast memory and no graphics card; it has its own section below.
On Windows, the software that drives the card also matters. Totsum runs NVIDIA cards through CUDA, and AMD and Intel cards through Vulkan. Two cards with the same amount of memory can differ in speed because of that path, not only because of the memory itself. And on some cards the app may not run at all, because the driver installed on the machine does not offer what the app needs, or because the maker no longer supports the card. The figures on this page are the makers’ own, from their specification sheets, and Totsum has not been tested on every card and chip named here; the trial, on your own machine, is the real test.
What each amount of memory lets you run
The sizes below are for the common compressed form of a model (4-bit), with room to work. They are rough, and they apply to any software that runs a model on your machine, not only to Totsum.
| Graphics memory on a PC, or unified memory on a Mac | What fits |
|---|---|
| 6 to 8 GB | Models of 3 to 8 billion parameters. Good for chat. For turning a question about a table into an exact query, a bigger model makes fewer mistakes. |
| 12 to 16 GB | Models of 12 to 14 billion parameters, with room to work. This is the class Totsum is built for. |
| 24 GB and more | Models of 24 to 32 billion parameters, or a 12 B model with a very long conversation kept in memory. |
| Mac, 18 GB or more | The same 12 B class. The system and your other applications share the pool, so 18 GB is the floor and 24 GB is comfortable. |
Three traps on a spec sheet
The laptop version of a card often has less memory than the desktop card with the same name. A desktop RTX 4070 has 12 GB. The laptop RTX 4070 has 8 GB. A desktop RTX 5070 has 12 GB; the laptop RTX 5070 exists with 8 GB and with 12 GB, and the shop listing does not always say which one is inside. On their makers’ sheets, the laptop RTX 4080 and the laptop RTX 5070 Ti both have 12 GB.
“AI PC” and “Copilot+” labels say nothing about graphics memory. They refer to a small extra chip (an NPU) made for light tasks. Totsum does not use it, so the GPU memory is still the figure to check.
System RAM is not graphics memory. A PC with 32 GB of RAM and integrated graphics can be fast and still not be compatible. On Windows, Totsum asks for 32 GB of RAM in addition to the graphics card.
Which Macs qualify
On a Mac, two numbers decide, and the chip’s name only hints at them.
The first is the amount of memory. It needs to be 18 GB or more, which is the floor the app checks when it starts, and a 16 GB Mac gets a warning. 24 GB is comfortable.
The second is how fast the chip reads memory, which decides how long an answer takes. It depends on the variant of the chip (base, Pro, Max, Ultra) far more than on its generation number. Apple’s figures, in GB/s, rounded:
| M1 | M2 | M3 | M4 | M5 | M6 | |
|---|---|---|---|---|---|---|
| Base chip | 70 | 100 | 100 | 120 | 153 | 170 |
| Pro | 200 | 200 | 150 | 273 | 307 | |
| Max | 400 | 400 | 300 to 400 | 410 to 546 | 460 to 614 | |
| Ultra | 800 | 800 | 800 | 1,200 |
The M6 (August 2026) exists as a base chip only for now; its 16 GB configuration reads memory at the M5’s speed.
The requirement on the hardware page draws its line at about 270 GB/s: an M4 Pro or an M5 Pro, or any Max or Ultra chip from the M3 generation up, with 18 GB or more. Those are the chip names that meet it. The table also shows what sits just below the line. An M3 Pro reads memory at half the speed of an M4 Pro, an M1 or M2 Pro at three quarters, so answers take noticeably longer there. An M1 or M2 Max reads it as fast as an M4 Pro on paper, with an older graphics part. And the base M5 and M6 are at a little over half the speed of an M4 Pro, where the base M1 to M4 are well below it. On any of those chips, with 18 GB or more, the app starts, and the trial tells you whether the speed suits you.
Each Mac model comes with its own range of chips. A MacBook Air always has a base chip. A 14-inch MacBook Pro comes with a base, a Pro or a Max chip, a 16-inch with a Pro or a Max. A Mac mini has a base or a Pro chip, a Mac Studio a Max or an Ultra, an iMac a base chip. About This Mac shows the exact name and the memory.
The new PCs with shared memory
A new family of Windows machines works like a Mac: one large pool of fast memory shared by the processor and the graphics part, with no separate graphics card. The first widely sold chip of this kind is the AMD Ryzen AI Max+ 395, found in laptops and mini PCs, with 32 GB to 128 GB of memory, of which a large share can be assigned to graphics in the machine’s settings.
On paper, such a machine has far more graphics memory than the requirement asks for, and AMD gives its memory speed as 256 GB/s, the level of a Pro chip in the Mac table above. The compatibility check reads the graphics memory the driver reports, which on these machines depends on that setting, and the app would drive the chip through Vulkan. Totsum has not been tested on one, so treat these machines as untested rather than unsupported, and run the trial before you rely on one.
Two machines that look similar are out of scope. NVIDIA’s DGX Spark ships with Linux only, and Totsum runs on Windows and macOS. And a laptop with an ordinary integrated graphics chip and 16 GB of RAM is not this family, however its box is labeled.
How to check your computer in one minute
On a Mac, open the Apple menu and choose About This Mac. The line Chip needs to say M4 Pro, M5 Pro, or a Max or Ultra chip of the M3 generation or later. The line Memory needs to say 18 GB or more.
On Windows, press Ctrl+Shift+Esc to open Task Manager, go to Performance and select your GPU. The line Dedicated GPU memory needs to say about 12 GB or more. Ignore the Shared GPU memory line, which is ordinary RAM. If the dedicated figure is small, a few hundred MB, the machine has no separate graphics card; if it says 8 GB, the card is below the requirement.
The full requirement, with example cards and Macs, is on the hardware page.
If your machine meets the requirement, you install Totsum, the installer downloads the model once (about 7 GB), and you open a file. You type a question about it in plain language, and a few minutes later you get a preview with a table or a chart, the rows behind it and the SQL query that produced it. The free trial gives you unlimited questions on the sample datasets and 10 on your own files, with no account and no time limit.
If your machine is below the line
There is no setting in the app that replaces missing memory. The model either fits with room to work, or it does not.
If your machine is close to the line, the free trial runs the real model on your hardware, so you can see how well it works before you pay.
If you are not sure which side of the line your machine is on, the compatibility check on the download page asks two questions and tells you, before you download anything.
What it costs you and what you get
Before you start, you spend some time. Checking that your machine is compatible takes a minute, and testing whether the speed suits you takes the free trial and an hour with a file of your own. A service in the cloud does not ask that of you. In return, the AI runs on your own machine, your file stays where it is on your disk, and you pay for the license once instead of paying a subscription.
Totsum Data Explorer is in a private beta, free in its first phase. You can apply for the beta; the free trial opens after it.