The best laptop for running an AI model offline is the one with the most memory you can get, because a model has to fit in memory before anything else matters. In our test an 8 GB Mac ran a 3B model only when few other apps were open, and a 4B model was already too slow. By our arithmetic 16 GB is the sensible floor, and it should fit models up to about 12B. Models around 20B to 27B want 24 or 32 GB.
We leave out prices and rankings, because both go stale fast. This page explains what decides whether a laptop can run a model, with our own measurements and the makers' own spec pages, so you can read any spec sheet yourself. We checked the specs on 25 October 2026.
What memory decides
A model runs from memory. Ollama, the program we used for our tests, pairs a 7.2 GB model download with a recommendation of 8 GB of available VRAM, or unified memory on a Mac. It adds that with less VRAM it can use system RAM, but responses may be slower.
The file size is only part of the bill. We measured Llama 3.2 3B at a 2.0 GB file and 2.5 GB of peak memory, and Gemma 3 4B at a 3.3 GB file and 3.5 GB of peak memory. On top of that, the operating system and a browser already take 4 to 5 GB. So we use this rough rule:
Memory needed = file size + about 0.5 GB + 4 to 5 GB for the system.
Long conversations cost more. Ollama says a larger context length increases the memory a model needs, and it picks the default from your VRAM (on a Mac, unified memory counts): 4k below 24 GiB, 32k at 24 to 48 GiB, and 256k at 48 GiB or more. Our guide to running AI offline on a Mac covers the setup itself.
What we measured on an 8 GB Mac
Our test machine was an Apple M2 with 8 GB. It was already using 7.5 GB of swap during the test, which is disk space standing in for missing memory. That makes it a harsh test, and an honest one, since most people have a browser open.
| Model | File on disk | Peak memory | First words | Speed | Load time |
|---|---|---|---|---|---|
| Llama 3.2 3B | 2.0 GB | 2.5 GB | 2.4 s | 29 per second | 13 s |
| Gemma 3 4B | 3.3 GB | 3.5 GB | 33 s (Mac swapping) | 8 per second | 42 s |
Gemma's 33 seconds came on a day when the Mac was short of memory. With little else running, Gemma answered in about 5 seconds, with the first words after 3 seconds. Llama stayed fast on both days, at about 4 seconds. That is the whole story of 8 GB: it works, until something else wants the memory.
The 1B models are small and fast. Gemma 3 1B used 1.0 GB at 72 per second, and Llama 3.2 1B used 1.5 GB at 57 per second. Neither was usable as an assistant, so shrinking the model further is not the answer.
Apple unified memory or a PC with a graphics card
On a Mac with Apple silicon, the processor and graphics share one pool of memory. Apple says this lets the entire chip reach a large single pool, which allows its Pro machines to run larger AI models completely on device. Ollama uses the Metal graphics interface on Apple devices. The MacBook Air page lists the Apple M5 chip with 16GB of unified memory, configurable to 24GB or 32GB. The MacBook Pro goes up to 128GB with the M5 Max.
Apple also lists memory bandwidth: 153GB/s for the M5 and 614GB/s for the top M5 Max. We have not measured how much that changes the speed of a model.
A PC works differently. The graphics card has its own memory, and a model runs best when it fits entirely there. Ollama says to avoid offloading the model to the CPU for best performance. NVIDIA's latest GeForce laptop chips list between 8 GB and 24 GB of graphics memory, and Ollama supports NVIDIA graphics cards with compute capability 5.0 or newer. So on a PC, read the graphics memory first and the system RAM second. A model whose file is 8.2 GB needs about 8.7 GB, so the next size up from 8 GB, which is 12 GB, is the one to look for.
Shared memory has reached PCs as well. AMD lists up to 128 GB of system memory on a 256-bit LPDDR5x bus for its Ryzen AI Max+ 395. Its page does not say how much of that the graphics side can use, and we did not test one.
How much RAM for a local LLM
The table below uses the rule from above with the file sizes from each model's Ollama page. Only the 8 GB row comes from our measurements. Every row above it is arithmetic, not a test.
| Memory | Left after system and browser | Models that should fit (file size) |
|---|---|---|
| 8 GB | 3 to 4 GB | Llama 3.2 3B (2.0GB). Gemma 3 4B (3.4GB) only just, as we saw. |
| 16 GB | 11 to 12 GB | Llama 3.1 8B (4.9GB), Gemma 3 12B (8.2GB) |
| 24 GB | 19 to 20 GB | gpt-oss 20B (14GB), Gemma 3 27B (17GB) with little room to spare |
| 32 GB | 27 to 28 GB | Gemma 3 27B (17GB) with room for a longer context |
| 64 GB | 59 to 60 GB | Llama 3.1 70B (43GB) |
Gemma 3 4B needs about 3.9 GB by our rule, against 3 to 4 GB left on an 8 GB laptop. That matches what we saw. One step further, gpt-oss 120B at 65GB needs about 65.5 GB for the model alone, so 64 GB is not enough for it. Apple's MacBook Pro page lists up to 128GB.
A bigger model is not automatically a better one for your job, and we have not tested the larger ones. What the arithmetic tells you is what can run at all, not how well it answers. Our own answers from a 3B model were wrong often enough that we would not trust a dose or a distance to it. See the results of our offline AI test.
Why storage matters for a library
Memory decides which model runs. Storage decides how much you can keep next to it. A text library is small: we estimate about 90 MB of text for 30,000 pages, with 2 to 5 GB more if you add pictures. The models themselves run from 2.0GB to 65GB on the Ollama pages above.
Big archives are another matter. The English Wikipedia file with pictures is 127.4 GB in drive-maker units, which the Kiwix listing shows as 119G. Our offline Wikipedia guide compares the smaller versions. Add a model such as gpt-oss 20B (14GB) and you want a drive with real room to spare. The MacBook Air page lists a 512GB SSD configurable to 1TB, 2TB or 4TB. Our offline survival library guide shows what is worth keeping and how big each part is.
A laptop that runs a model in a blackout still needs power. Our laptop runtime tool estimates how long a power station will keep it going.
What we are targeting, and what we have not tested
We're building Waystone, an offline preparedness assistant that is still in development and not on sale. First target: recent Mac laptops with Apple silicon and 16 GB of memory. Other computers are being considered. Exact requirements will be confirmed after testing. We chose 16 GB because of the 8 GB result above: the 3B model needs few other apps open, and a 4B model was already too slow.
There are gaps in what we know:
- We did not test a 16 GB Mac. Its row in the table is arithmetic.
- Windows was not tested, and we have not run any laptop with an NVIDIA or AMD chip.
- The memory figure for a 7B to 8B model (about 5.5 to 6.5 GB) is an estimate, not a measurement.
- The Ollama and Apple pages change. Check them again before you buy.