How Much RAM Do You Need to Run Ollama? (2026 Buying Guide)
This is the single most-asked question about running AI at home, and it has a precise answer — unlike most tech buying advice. Here it is, straight, with the reasoning so you can apply it to anything you already own or are about to buy.
Want the full hardware decision made for you? Private AI at Home includes a 10-question hardware picker checklist. $19, free lifetime updates.
Why RAM (not CPU, not “AI chips”) is the answer
An AI model is a file that has to be loaded into fast memory to run. If it fits, it runs — at a speed roughly proportional to how fast that memory is. If it doesn’t fit, it either fails to load or crawls at a punishing swap-to-disk speed. CPU generation, core count and “AI TOPS” marketing barely move the outcome once you clear this bar. Memory is the whole game.
The table
| RAM available | Model class that fits | Real-world feel |
|---|---|---|
| 8 GB | 1–3B | Barely usable — fine for testing, not daily use |
| 16 GB | 3–4B | A genuinely useful quick helper: drafts, summaries, simple Q&A |
| 32 GB | 7–8B | The sweet spot. A daily-driver assistant that rarely feels limited |
| 64 GB | 13–14B | Noticeably sharper reasoning, writing, and multi-step tasks |
| 96 GB+ | 27–32B | “This runs in my house?” territory — approaches recent flagship cloud models |
Two notes that change the math:
- Apple Silicon Macs use unified memory — the number on the spec sheet (16, 24, 32 GB) is the number in this table directly. A 16 GB M-series Mac genuinely runs the 7–8B row well, punching above the same RAM figure on Windows.
- Dedicated GPU VRAM beats the same amount of system RAM for speed (not for what fits) — an 8 GB graphics card runs the 7–8B row faster than 32 GB of plain system RAM running the same model, even though both technically “fit” it. RAM decides what runs; VRAM decides how fast.
Match RAM to what you’ll actually do
- Quick drafts, casual questions, kids’ homework help: 16 GB is genuinely enough. Don’t overbuy.
- A real daily assistant — writing, summarizing, document Q&A, brainstorming: aim for 32 GB. This is where most households land and stay happy.
- Heavier reasoning, coding help, longer documents: 64 GB stops feeling like a compromise.
- You want to see what the fuss is about with the biggest home-feasible models: 96 GB+, usually via a Mac Studio or a multi-GPU build — a smaller, more committed audience.
Check what you already own before buying anything
Open Ollama on your current laptop or desktop and try ollama run llama3.2 (a small 3B model, ~2 GB download). If you have 16 GB total system RAM, you likely already have a usable setup — the actual buying decision only starts once you know whether 16 GB felt limiting for what you wanted to do.
What RAM alone won’t tell you
RAM decides what can run, not the full experience. Speed also depends on whether it’s VRAM or plain RAM (see note above), and getting a model to feel “yours” — a proper chat interface, family accounts, your own documents wired in, safe access from your phone away from home — is a separate, non-technical set of steps most RAM guides skip entirely. I wrote the full path once the hardware question is settled:
Private AI at Home — the non-techie's playbook →
$19 · model cheatsheet for every RAM tier included · free lifetime updates.