Ollama commands cheat sheet: every command you'll actually use

Ollama has about a dozen commands, and you'll use five of them every day: run, pull, list, ps and rm. Here's the full set, with what each one is for. If you haven't installed Ollama yet, start with your first model in ten minutes.

The everyday commands

Command	What it does
`ollama run qwen3:8b`	Download the model if needed, then chat with it
`ollama pull qwen3:8b`	Download or update a model without starting a chat
`ollama list`	Show every model on disk, with its size
`ollama ps`	Show loaded models and whether they're on the GPU
`ollama stop qwen3:8b`	Unload a model from memory now
`ollama rm qwen3:8b`	Delete a model from disk
`ollama show qwen3:8b`	Show a model's details: size, quant, context length
`ollama cp qwen3:8b my-model`	Copy a model under a new name
`ollama create my-model -f Modelfile`	Build a custom model from a Modelfile
`ollama serve`	Start the server by hand (the app normally does this)
`ollama --version`	Show which version you have

Run ollama help, or add --help to any command, for the full options.

Running models

Ask a single question without opening a chat:

ollama run qwen3:8b "Explain what VRAM is in two sentences"

See how fast it's going. --verbose prints timing after each answer, including the eval rate in tokens per second:

ollama run qwen3:8b --verbose

Compare that number with our calculator's estimate for your card on every GPU. If yours is far lower, check ollama ps.

Checking the GPU

ollama ps

The PROCESSOR column is the most useful thing in Ollama. 100% GPU is what you want. A split like 30%/70% CPU/GPU means the model didn't fit in graphics memory and part of it runs on the processor, which is several times slower. Pick a smaller model or read how much VRAM you need.

Inside a chat

Once you see the >>> prompt, these slash commands work:

/bye quits the chat.
/? lists every slash command.
/clear wipes the conversation so far.
/show info shows the model's details.
/set system "Answer briefly." sets a system prompt for this chat.
/set parameter num_ctx 8192 raises the context length for this chat. More context uses more memory.

To paste several lines at once, start and end the text with three double quotes (""").

Model tags and sizes

The part after the colon is the tag. It picks the size, and sometimes the quant. With no tag you get latest, which is usually the default size. Here are the download sizes for the models we track:

Tag	Download	Fits on
`qwen3:4b`	2.5 GB	Almost anything, including CPU only
`qwen3:8b`	5.2 GB	8 GB cards
`gemma3:12b`	8.1 GB	12 GB cards
`qwen3:14b`	9.3 GB	12 GB cards
`gpt-oss:20b`	14 GB	16 GB cards
`qwen3:30b`	19 GB	24 GB cards
`qwen3:32b`	20 GB	24 GB cards
`llama3.3:70b`	43 GB	Two 24 GB cards, or a 64 GB Mac

Each model has a page with speeds on every GPU, for example Qwen3 14B or Qwen3 30B-A3B. Browse them all on models.

Settings worth knowing

Ollama is configured with environment variables, not a settings file. The ones people use most:

OLLAMA_HOST sets the address the server listens on. The default is 127.0.0.1:11434, this computer only.
OLLAMA_MODELS moves where models are stored. See where Ollama stores models.
OLLAMA_KEEP_ALIVE sets how long a model stays loaded after the last request. The default is five minutes.

How you set them depends on your system. The Windows, Linux and Mac guides each show how.

The API

Everything the commands do is also available over HTTP at http://localhost:11434. A quick test:

curl http://localhost:11434/api/generate -d '{"model": "qwen3:8b", "prompt": "Say hello", "stream": false}'

That's how chat apps like Open WebUI and coding tools talk to Ollama.

Where to go next

Running out of disk? Read how to update and delete models.
Not sure what the words mean? The local AI glossary covers tokens, quants and context.
Want something to do with it? Try our first week with local AI.