Ollama commands: the cheat sheet

Ollama commands are few and easy to remember: ollama run starts a model, ollama pull downloads one, ollama ls lists what you have, ollama ps shows what is loaded, ollama stop unloads it and ollama rm deletes it. This cheat sheet lists every command in Ollama's CLI reference, the environment variables from its FAQ and the addresses of its local API, as documented on October 6, 2026. If you have not installed Ollama yet, start with how to run an LLM locally.

Which Ollama commands do you need most?

Table: the Ollama command line (from Ollama's CLI reference)

The reference gives gemma4 as its example model; any model name from the Ollama library works in the same way, for instance qwen3:8b. For the full list of options on your installed version, run ollama --help.

How do you run, download and remove a model?

ollama run gemma4 starts a chat with the model. ollama pull gemma4 only downloads it, which is useful before you go offline. ollama ls shows what you have on disk so you can decide what to delete, and ollama rm gemma4 removes a model and frees the space. Which model to pick depends on your memory; our VRAM calculator and GPU guide show what fits.

How do you see what is running, and free the memory?

ollama ps lists the loaded models and, in the PROCESSOR column, whether they run on the GPU: "100% GPU" means the whole model is in video memory, and any other split means part of it runs on the CPU, which is slower. By default a model stays loaded for five minutes after its last use. To unload it at once, run ollama stop followed by the model name. To keep a model loaded, set keep_alive to a negative number such as -1 in the API or with the OLLAMA_KEEP_ALIVE variable, and to unload it immediately after a request set keep_alive to 0.

Which environment variables does Ollama use?

Set a variable before you start the server, for example OLLAMA_CONTEXT_LENGTH=64000 ollama serve. Ollama's FAQ lists these:

Table: Ollama environment variables (from the FAQ)

Lowering the cache type from f16 to q8_0 roughly halves the memory the cache uses, which can let a longer context fit.

Where does Ollama store its models?

On macOS the models are in ~/.ollama/models, on Linux in /usr/share/ollama/.ollama/models and on Windows in C:\Users\%username%\.ollama\models. Set OLLAMA_MODELS to a different folder if your system drive is small.

How do you use the Ollama API from other programs?

When the server runs, Ollama listens on http://localhost:11434. Its native API lives under http://localhost:11434/api. Ollama's documentation shows a chat request for its cloud service and says you can use it locally by replacing the URL with http://localhost:11434/api/chat and removing the authorization header, which gives a request like this:

curl http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{"model": "gemma4", "messages": [{"role": "user", "content": "Why is the sky blue?"}], "stream": false}'

Tools written for other APIs can use Ollama too. For OpenAI-compatible clients, the base URL is http://localhost:11434/v1/ and the API key is a placeholder (ollama) that Ollama ignores; it supports chat completions, completions, models, embeddings and responses endpoints. For Anthropic-compatible clients, the base URL is http://localhost:11434 with the key ollama, and for Claude Code the documentation sets ANTHROPIC_AUTH_TOKEN=ollama and ANTHROPIC_BASE_URL=http://localhost:11434. Ollama supports a subset of each API, so check the compatibility pages if a feature is missing.

Can Ollama create embeddings and launch other tools?

Yes. Embedding models run with the same command: ollama run embeddinggemma "Hello world" or, with a pipe, echo "Hello world" | ollama run nomic-embed-text. Embeddings are the building block of retrieval-augmented generation. And ollama launch connects other applications to Ollama models: ollama launch claude starts Claude Code, ollama launch claude --model qwen3.5 chooses the model, and ollama launch droid --config configures an integration without starting it.

How do you fix common problems with these commands?

• Answers are slow: run ollama ps. A split between CPU and GPU means the model does not fit; use a smaller model or a more compressed one.
• The computer runs out of memory: run ollama stop on models you are not using, or lower OLLAMA_MAX_LOADED_MODELS and the context length.
• The disk is full: run ollama ls, then ollama rm on the models you no longer use.
• A program cannot connect: check that ollama serve is running and that the program points to 127.0.0.1:11434, or to the address you set in OLLAMA_HOST.
• You want to compare runners: see Ollama vs LM Studio.