Local AI model suggestions — The Privacy Diet

By memory

Find the row for your computer, and pick the largest model that fits. The book's own suggestions come first; newer releases are beside them. Sizes are the download from Ollama's library — the model needs a couple of gigabytes more than that to run (Chapter 11).

Your memory The book suggests Newer, worth trying

Your memory	The book suggests	Newer, worth trying
8 GB	`qwen3:4b` · 2.5 GB	`qwen3.5:4b` · 3.3–4.0 GB · reads images too
16 GB	`qwen3:8b` · 5.2 GB `gpt-oss:20b` · 14 GB · close other apps first	`qwen3.5:9b` · 6.6–7.6 GB · reads images too `gemma4:12b` · 7.7–8.0 GB · reads images too
32 GB, or a 24 GB graphics card	The 20–30B versions of current Qwen, Gemma or gpt-oss models	`qwen3.5:27b` · 17–20 GB · reads images too `gemma4:26b` · 16–19 GB · reads images too

8 GB

qwen3:4b · 2.5 GB

qwen3.5:4b · 3.3–4.0 GB · reads images too

16 GB

qwen3:8b · 5.2 GB

gpt-oss:20b · 14 GB · close other apps first

qwen3.5:9b · 6.6–7.6 GB · reads images too

gemma4:12b · 7.7–8.0 GB · reads images too

32 GB, or a 24 GB graphics card

The 20–30B versions of current Qwen, Gemma or gpt-oss models

qwen3.5:27b · 17–20 GB · reads images too

gemma4:26b · 16–19 GB · reads images too

Getting one

On your computer, with Ollama (Chapter 12):

ollama run qwen3.5:9b

On the household server (Chapter 13):

docker exec -it ollama ollama pull qwen3.5:9b

Swap in any name from the table. In LM Studio, search for the family name — Qwen3.5, Gemma 4, gpt-oss — and it will mark which versions fit your computer.

Specialists

Reading images and scanned letters: every size of Qwen3.5 and Gemma 4 can take an image as well as text.
Code: qwen2.5-coder:7b, the model from Be Your Own Cloud, Chapter 16, or qwen2.5-coder:14b with 16 GB or more of graphics memory.
Searching your own documents: keep nomic-embed-text, the embedding model Chapters 13 and 15 set up. Changing it means preparing your documents again.

Changes