Pull and Run Models

The model:tag namespace

Every Ollama model is addressed as name:tag. The default tag (whatever ollama pull qwen2.5 resolves to) is a sensible quantization picked by the model maintainers. To pin a specific quantization or size, use the explicit tag form.

Example: qwen2.5:7b-instruct-q5_K_M reads as:

qwen2.5 — model family
7b — parameter size
instruct — variant (instruction-tuned, not base)
q5_K_M — quantization

Pulling and running

ollama pull downloads the model blobs (layered like Docker layers — same blob shared across tags is downloaded once). ollama run starts an interactive chat REPL and pulls the model first if you don't have it. Both commands resume on partial downloads.

Where models live

On macOS: ~/.ollama/models/. On Linux: /usr/share/ollama/.ollama/models/. Move them to an external drive with the OLLAMA_MODELS environment variable — useful when your boot drive is small.

EV v3 · 2026-09-18: The relocate-your-model-store snippet used to export OLLAMA_MODELS, launchctl setenv it, and pull in the next breath. That could not work and never could: the daemon reads its environment when it starts, so the pull went to the already-running daemon and landed in the old store, while export only ever touched the shell you typed it in. Ollama's own FAQ has the missing step — set the variable, then restart Ollama — and the block now does set, restart, verify, pull, in that order. The restart line itself needed a second pass, caught in review: the first repair put the desktop app's osascript form on the live line, but the macOS path the install lesson teaches is Homebrew, where that command fails and && then skips the relaunch entirely — leaving the next pull aimed at the old store again. The Homebrew line is live now and the others are comments.

Code

Pull, run, list, inspect·bash

# Pull the default tag
ollama pull qwen2.5

# Pull a specific quant
ollama pull qwen2.5:7b-instruct-q5_K_M

# Run interactively (pulls if missing)
ollama run qwen2.5
# /bye to exit, /? for help

# List installed models with sizes
ollama list

# Inspect a specific model
ollama show qwen2.5:7b
ollama show qwen2.5:7b --modelfile

# Currently loaded models + memory usage
ollama ps

Move models to external storage·bash

# The daemon reads this at startup, so the order matters:
# set it, restart Ollama, and only then pull.

# 1. Persist it where the daemon will see it (macOS / launchd)
launchctl setenv OLLAMA_MODELS /Volumes/ExternalSSD/ollama-models
#    Linux (systemd):
#      sudo systemctl edit ollama.service
#      Environment="OLLAMA_MODELS=/mnt/ssd/ollama-models"

# 2. Restart the daemon so it picks the new store up. Use the line that
#    matches how you installed it — the install lesson's macOS path is
#    Homebrew, and the osascript form only works for the desktop app.
brew services restart ollama                    # Homebrew (macOS)
# osascript -e 'tell application "Ollama" to quit'; open -a Ollama   # app
# sudo systemctl daemon-reload && sudo systemctl restart ollama       # Linux

# 3. Confirm the store moved, THEN re-pull
ollama list                 # expect an empty list at the new location
ollama pull qwen2.5:7b

# `export` alone only affects the shell you typed it in — the daemon
# is a separate, already-running process and never sees it.

Exercise

Pull two flavors of the same model: the default tag and an explicit larger quantization. Run ollama list and ollama show on both. Note the size difference, and the context length each one reports. That number is the architecture's maximum, not the window a single request gets: Ollama's runtime default is far smaller unless the model or your options raise it, which is why num_ctx has its own lesson in the API track.