Why local AI matters

Free
0/6 steps visited

Step 3 of 6

Pick a starter model for your RAM

Read · ~6 min

Pick a starter model for your RAM

As of June 2026 — model names on ollama.com/library change over time. The sizes below are the default Q4_K_M downloads Ollama uses (a good balance of quality and disk space).

You're not choosing "the best AI in the world" today. You're choosing what your machine can run comfortably for a first win. Bigger models come later, once you know the ropes.

How much RAM do you have?

If you have…Pull this modelDownload (approx)Why
8 GB total RAMllama3.2~2.0 GBMeta's small 3B model — fast enough to learn on, fits tight machines
16 GB RAMllama3.1:8b or qwen2.5:7b~4.7–4.9 GBA solid daily driver; noticeably smarter than 3B
24 GB+ RAMqwen2.5:14b or mistral-nemo~7.1–9.0 GBStronger reasoning; still one-laptop friendly

Not sure how much RAM you've got? Check your system:

  • Windows: Settings → System → About → Installed RAM
  • macOS: Apple menu → About This MacMemory
  • Linux: run free -h in a terminal and read the total line

Optional swaps (same RAM tier)

  • 8 GB and want slightly sharper answers: try phi4-mini (~2.5 GB) instead of llama3.2.
  • 16 GB and you'll write a lot of code later: qwen2.5:7b follows instructions especially well.

Write this down

You'll need it on the next pages. Something like:

My RAM: 16 GB
My model: llama3.1:8b
Pull command: ollama pull llama3.1:8b
Run command: ollama run llama3.1:8b

Take it from me — I once burned a whole evening pulling a 70B model onto a laptop that simply couldn't breathe. Don't be me. Match the model to the machine; you can always pull a bigger one once you know what you're doing.

Got a dedicated GPU? Lovely — replies come back faster. No GPU? Also fine. Ollama happily uses your CPU and system RAM. There's no gatekeeping here on "you need a £2k gaming PC."

Continue — what Off the Grid will prove.

Warning

Real power. Educational use only.

What we teach you to build is genuinely powerful — uncensored assistants, agents, and automations on your own hardware. In the wrong hands, that is as dangerous as malicious code in the wrong hands. We do not teach illegal, malicious, or harmful use. You are responsible for what you deploy.

See what we mean →