Inside The Machine
00Guide · 4 minutes to read

Mac or PC for local AI?

There is no one winner. Memory decides which AI models fit. Then speed decides how it feels.

Prices and speeds: machines.json, checked 2026-10-15. Model sizes: models.json, 2026-10-26.

1Four kinds

Four kinds of computer can run AI.

The site sorts every computer into one of these four.

  1. 1
    Mac

    One box. Memory is shared by the chip and the graphics.

  2. 2
    PC with a graphics card

    The model runs in the card's own memory. Prices are for the card alone.

  3. 3
    Mini PC or AI box

    A small PC with lots of shared memory, built for AI.

  4. 4
    Small board

    Tiny, cheap and low power. Slow for chat.

"Shared memory" means one pool of memory for the main chip and the graphics. The AI can use most of it. A graphics card has its own memory, called VRAM (video memory). On a PC with a card, the model runs in the VRAM.

2Memory first

Rule 1: the model must fit in memory.

Too big, and it is very slow or does not start.

  • On a Mac, we count about 75% of the memory as usable for AI. The rest is for the system and your apps. (This is the macOS default.)
  • On a PC with a graphics card, the card's VRAM counts, not the computer's memory. We count all of it but 1 GB.
  • A mini PC or AI box shares its memory like a Mac. We use the share its maker gives the graphics.
  • A small board is cheap and uses little power, but it is slow for chat.

Example: the open model gpt-oss-120b needs about 74.2 GB of memory. Its little brother gpt-oss-20b needs about 15.4 GB. Here is which computer fits the big one.

Does gpt-oss-120b (needs about 74.2 GB) fit? One example per kind.
ComputerKindMemoryUsable for AIFits gpt-oss-120b?Price
Mac Studio (M5 Max)Mac128 GB96 GBYes$5,099 (worked out from list prices, macworld.com, 2026-10-15)
GeForce RTX 5090PC with a graphics card32 GB31 GBNo$5,700 (lowest new price in stock, rtx.live, 2026-10-15), card only
Framework Desktop (Ryzen AI Max+ 395)Mini PC or AI box128 GB96 GBYes$3,449 (list price, computingforgeeks.com, 2026-10-15), DIY kit, out of stock
Raspberry Pi 5 (16GB), CPU onlySmall board16 GB12 GBNo$305 (list price, computingforgeeks.com, 2026-10-12), board only

So the GeForce RTX 5090 is a very fast card, but its 32 GB is too small for the big model. The Mac Studio (M5 Max) with 128 GB holds it.

3Then speed

Rule 2: then speed decides how it feels.

Speed is counted in tokens per second (tok/s). A token is a piece of a word.

What sets speed is memory bandwidth: how fast the computer can read the model. A graphics card reads its VRAM very fast. Published tests:

  • Mac Studio (M5 Max): 78.2 tok/s on gpt-oss-120b (oMLX v0.3.9; test machine: M5 Max 40c, 128GB). Source: omlx.ai, 2026-10-15.
  • GeForce RTX 5090: 318.4 tok/s on gpt-oss-20b (llama.cpp; test machine: RTX 5090 + Ryzen 7 9850X3D (chat-short workload)). Source: llm-speed.com, 2026-10-15.
  • Framework Desktop (Ryzen AI Max+ 395): 54 tok/s on gpt-oss-120b (llama.cpp Vulkan; test machine: Framework Desktop 128GB (llama-bench pp512/tg128)). Source: llamaperf.com, 2026-10-15.
  • Raspberry Pi 5 (16GB), CPU only: 5.6 tok/s on llama3.2 3B (Ollama; test machine: Pi 5 8GB CPU (Jeff Geerling)). Source: computingforgeeks.com, 2026-10-12.

Bandwidth in our data: Mac Studio (M5 Max) 614 GB/s, GeForce RTX 5090 1792 GB/s, Framework Desktop (Ryzen AI Max+ 395) 256 GB/s, Raspberry Pi 5 (16GB), CPU only 17 GB/s. Tests use different engines and settings, so compare them loosely. What a token is: What is a token?

4Choose

So which one?

Start from what you want to run, not from the brand.

  1. 1
    You have a computer already

    Use it first. Scan my machine reads your memory, chip and graphics card with one read-only command and lists what fits.

  2. 2
    You want big models

    Look for lots of memory: a Mac or a mini PC with shared memory. The Machines page lists each one with its price and source.

  3. 3
    You want speed on small models

    A PC with a strong graphics card is very fast, as long as the model fits in its VRAM. Card prices are for the card alone: you also need a PC.

  4. 4
    Windows or Mac steps

    Our recipes are tested on a Mac first. The Windows steps are there, but we have not tested them yet.

Next

Keep going.

Every guide is on the Learn page.

Compare machines Scan my machineWhat is a token?
Agent view · text/markdown · or send Accept: text/markdown

Mac or PC for local AI? A plain-words guide 


URL: https://insidethemachine.ai/learn/mac-or-pc
Markdown: https://insidethemachine.ai/learn/mac-or-pc.md
Read: 4 minutes
Data: Prices and speeds: machines.json, checked 2026-10-15. Model sizes: models.json, 2026-10-26.

There is no one winner. Memory decides which AI models fit. Then speed decides how it feels.

1. Four kinds of computer can run AI. 


The site sorts every computer into one of these four.

1. **Mac**: One box. Memory is shared by the chip and the graphics.
2. **PC with a graphics card**: The model runs in the card's own memory. Prices are for the card alone.
3. **Mini PC or AI box**: A small PC with lots of shared memory, built for AI.
4. **Small board**: Tiny, cheap and low power. Slow for chat.

"Shared memory" means one pool of memory for the main chip and the graphics. The AI can use most of it. A graphics card has its own memory, called VRAM (video memory). On a PC with a card, the model runs in the VRAM.

2. Rule 1: the model must fit in memory. 


Too big, and it is very slow or does not start.

- On a Mac, we count about 75% of the memory as usable for AI. The rest is for the system and your apps. (This is the macOS default.)
- On a PC with a graphics card, the card's VRAM counts, not the computer's memory. We count all of it but 1 GB.
- A mini PC or AI box shares its memory like a Mac. We use the share its maker gives the graphics.
- A small board is cheap and uses little power, but it is slow for chat.

Example: the open model gpt-oss-120b needs about 74.2 GB of memory. Its little brother gpt-oss-20b needs about 15.4 GB. Here is which computer fits the big one.

Does gpt-oss-120b (needs about 74.2 GB) fit? One example per kind.
| Computer | Kind | Memory | Usable for AI | Fits gpt-oss-120b? | Price |
|---|---|---|---|---|---|
| Mac Studio (M5 Max) | Mac | 128 GB | 96 GB | Yes | $5,099 (worked out from list prices, macworld.com, 2026-10-15) |
| GeForce RTX 5090 | PC with a graphics card | 32 GB | 31 GB | No | $5,700 (lowest new price in stock, rtx.live, 2026-10-15), card only |
| Framework Desktop (Ryzen AI Max+ 395) | Mini PC or AI box | 128 GB | 96 GB | Yes | $3,449 (list price, computingforgeeks.com, 2026-10-15), DIY kit, out of stock |
| Raspberry Pi 5 (16GB), CPU only | Small board | 16 GB | 12 GB | No | $305 (list price, computingforgeeks.com, 2026-10-12), board only | 

So the GeForce RTX 5090 is a very fast card, but its 32 GB is too small for the big model. The Mac Studio (M5 Max) with 128 GB holds it.

3. Rule 2: then speed decides how it feels. 


Speed is counted in tokens per second (tok/s). A token is a piece of a word.

What sets speed is memory bandwidth: how fast the computer can read the model. A graphics card reads its VRAM very fast. Published tests:

- Mac Studio (M5 Max): **78.2 tok/s** on gpt-oss-120b (oMLX v0.3.9; test machine: M5 Max 40c, 128GB). Source: [omlx.ai](https://omlx.ai/benchmarks/dgadv0k9), 2026-10-15.
- GeForce RTX 5090: **318.4 tok/s** on gpt-oss-20b (llama.cpp; test machine: RTX 5090 + Ryzen 7 9850X3D (chat-short workload)). Source: [llm-speed.com](https://llm-speed.com/m/gpt-oss-20b), 2026-10-15.
- Framework Desktop (Ryzen AI Max+ 395): **54 tok/s** on gpt-oss-120b (llama.cpp Vulkan; test machine: Framework Desktop 128GB (llama-bench pp512/tg128)). Source: [llamaperf.com](https://llamaperf.com/model/gpt-oss), 2026-10-15.
- Raspberry Pi 5 (16GB), CPU only: **5.6 tok/s** on llama3.2 3B (Ollama; test machine: Pi 5 8GB CPU (Jeff Geerling)). Source: [computingforgeeks.com](https://computingforgeeks.com/edge-llm-jetson-orin-vs-ai-hat-2-vs-hailo-10h/), 2026-10-12.

Bandwidth in our data: Mac Studio (M5 Max) 614 GB/s, GeForce RTX 5090 1792 GB/s, Framework Desktop (Ryzen AI Max+ 395) 256 GB/s, Raspberry Pi 5 (16GB), CPU only 17 GB/s. Tests use different engines and settings, so compare them loosely. What a token is: [What is a token?](https://insidethemachine.ai/learn/tokens)

4. So which one? 


Start from what you want to run, not from the brand.

1. **You have a computer already**: Use it first. [Scan my machine](https://insidethemachine.ai/my) reads your memory, chip and graphics card with one read-only command and lists what fits.
2. **You want big models**: Look for lots of memory: a Mac or a mini PC with shared memory. The [Machines](https://insidethemachine.ai/machines) page lists each one with its price and source.
3. **You want speed on small models**: A PC with a strong graphics card is very fast, as long as the model fits in its VRAM. Card prices are for the card alone: you also need a PC.
4. **Windows or Mac steps**: Our recipes are tested on a Mac first. The Windows steps are there, but we have not tested them yet.

Next 


- [Compare machines](https://insidethemachine.ai/machines) (markdown: https://insidethemachine.ai/machines.md)
- [Scan my machine](https://insidethemachine.ai/my) (markdown: https://insidethemachine.ai/my.md)
- [What is a token?](https://insidethemachine.ai/learn/tokens) (markdown: https://insidethemachine.ai/learn/tokens.md)

More guides: https://insidethemachine.ai/learn.md