Skip to content
SizeMyLaptop

How much RAM does a local LLM need?

A local model has to fit in memory. Ollama lists the file size of each model, and LM Studio states a RAM recommendation for its app. Context and cache come on top of the file.

Figures read from the vendors' own pages on 2026-10-20.

What the vendors say

Vendor minimum and recommended RAM for Local LLMs4 GB8 GB16 GBLM Studio (macOS)16LM Studio (Windows)16
Vendor minimumVendor recommendation Gridlines are the RAM sizes Macs come in.
File size of each local model, as Ollama lists itllama3.2:1b1.3 GBqwen3:1.7b1.4 GBgemma3:4b3.4 GBmistral:7b4.4 GBllama3.1:8b4.9 GBgemma3:12b8.2 GBdeepseek-r1:14b9 GBphi4:14b9.1 GBqwen3:14b9.3 GBgpt-oss:20b14 GBgemma3:27b17 GBdeepseek-r1:32b20 GBllama3.1:70b43 GBgpt-oss:120b65 GB
File size of each model, as Ollama lists it. Context and cache come on top.
Show the sizes as a table, with the quoted lines
Model file sizes listed by Ollama
ModelFile sizeSource
llama3.2:1b1.3 GBOllama
Quoted text
llama3.2:1b baf6a787fdff • 1.3GB
qwen3:1.7b1.4 GBOllama
Quoted text
qwen3:1.7b 8f68893c685c • 1.4GB
gemma3:4b3.4 GBOllama
Quoted text
gemma3:4b latest 68ceab7e468a • 3.4GB
mistral:7b4.4 GBOllama
Quoted text
mistral:7b latest 6577803aa9a0 • 4.4GB
llama3.1:8b4.9 GBOllama
Quoted text
llama3.1:8b latest 46e0c10c039e • 4.9GB
gemma3:12b8.2 GBOllama
Quoted text
gemma3:12b b92b218582c5 • 8.2GB
deepseek-r1:14b9 GBOllama
Quoted text
deepseek-r1:14b c333b7232bdb • 9.0GB
phi4:14b9.1 GBOllama
Quoted text
phi4:14b latest ac896e5b8b34 • 9.1GB
qwen3:14b9.3 GBOllama
Quoted text
qwen3:14b bdbd181c33f2 • 9.3GB
gpt-oss:20b14 GBOllama
Quoted text
gpt-oss:20b latest f38aa0c53da5 • 14GB
gemma3:27b17 GBOllama
Quoted text
gemma3:27b 6be2e2ba6843 • 17GB
deepseek-r1:32b20 GBOllama
Quoted text
deepseek-r1:32b edba8017331d • 20GB
llama3.1:70b43 GBOllama
Quoted text
llama3.1:70b 711a9e8463af • 43GB
gpt-oss:120b65 GBOllama
Quoted text
gpt-oss:120b 714d86958b6e • 65GB

What it does with memory

A local model loads its weights into memory, and a longer conversation adds context on top. On Apple silicon the weights share unified memory with every other app.

Signs it is short

  1. 1

    Answers arrive far slower than they started

  2. 2

    The machine freezes when the model loads

  3. 3

    Swap grows during generation

The file size is a floor. Compare it with the memory you have left after your other apps. How to read memory pressure.

What more RAM costs

Extra cost per GB of each RAM step, MacBook Air 13-inch, Apple NLMacBook Air 13-inch, M5 Chip: 16 to 24 GB€27,50
Extra cost per GB of each step, in euros, VAT included.

Sources, read 2026-10-20: Apple NL: 16 GB page and 24 GB page; Apple US: 16 GB page and 24 GB page. All steps.

Try these figures in the calculator.

Open the calculator with this workload

Then measure it for real: get SizeMyLaptop and see your own memory pressure over a working fortnight.

Get SizeMyLaptop

  1. Open Terminal

    Press Command and Space, type Terminal, press Return.

  2. Paste it and press Return

    One command installs the tool, checks its checksum, then starts recording. No password, no admin rights.

    curl -fsSL https://sizemylaptop.com/install.sh | bash -s -- --start

    Rather read it first? Download it, read it, then run it:

    curl -fsSLO https://sizemylaptop.com/install.sh && less install.sh && bash install.sh --start

  3. Done. Keep working as usual.

    After one to two weeks, run baseline report in Terminal (the installer prints the exact command). It stops recording by itself after 14 days and the report opens in your browser.

Free. Nothing leaves your Mac unless you click Share. No password, no admin rights. Needs Python 3 (a Mac without it shows Apple's one-time install window); installs nothing else.