Ollama Cheat Sheet

1Install & first runone command

curl -fsSL https://ollama.com/install.sh | sh # Linux★
macOS & Windows have native installers on ollama.com. Installing starts the background server automatically.
ollama run llama3.2★
Pulls the model on first use, then opens an interactive chat. Type a message; /bye to exit, /? for in-chat commands.
ollama run llama3.2 "Explain PagedAttention in one line"
Pass a prompt inline for a one-shot answer (great for scripts / pipes).
echo "summarize this" | ollama run llama3.2
Reads stdin — compose it with other shell tools.

2Core CLI commandsdaily driver

ollama list # alias: ollama ls★
Show downloaded models with size + last-modified.
ollama ps★
Show models currently loaded in memory and their VRAM/CPU split.
ollama stop llama3.2 · ollama rm llama3.2
stop unloads a running model; rm deletes it from disk.
ollama serve
Run the server in the foreground (normally it's already running). Reads config from environment variables (card 11).
ollama --version · ollama show llama3.2
show prints a model's parameters, template, license, and Modelfile.

3Model managementpull & share

ollama pull llama3.2:3b # tag = size/variant★
Tags select size/quant: llama3.2:1b, :3b, qwen3:8b, :latest. Browse at ollama.com/library.
ollama run hf.co/<user>/<repo>-GGUF:Q4_K_M
Run any GGUF model straight from Hugging Face by prefixing hf.co/ and picking a quantization tag.
ollama cp llama3.2 my-llama · ollama push user/model
cp clones a model locally; push uploads to your ollama.com namespace (after ollama signin).
# a 70B model needs ~40GB+ RAM/VRAM even quantizedgotcha
Pick a size your hardware can hold, or Ollama offloads to CPU and slows down. ollama ps shows the split.

5/api/generate & /api/chatlocalhost:11434

curl localhost:11434/api/chat -d '{ "model":"llama3.2", "messages":[{"role":"user","content":"Hi"}] }'★
/api/chat = multi-turn with message history. The response's message.content holds the reply.
curl localhost:11434/api/generate -d '{ "model":"llama3.2","prompt":"Why is the sky blue?" }'★
/api/generate = single-turn completion from a raw prompt.
"stream": falsenote
Both endpoints stream by default (NDJSON, one JSON object per token). Set "stream": false for a single response object.
"options": {"temperature": 0.2, "num_ctx": 4096}
Per-request overrides go in options. Other endpoints: /api/tags (list), /api/ps, /api/pull, /api/show.

6Embeddingsvectors

curl localhost:11434/api/embed -d '{ "model":"nomic-embed-text","input":"hello" }'★
/api/embed returns embeddings under embeddings. input accepts a string or an array for batching.
ollama pull nomic-embed-text # or mxbai-embed-large, bge-*
Dedicated embedding models are small & fast — ideal for local RAG.

7chat / generatepip install ollama

from ollama import chat r = chat(model="llama3.2", messages=[{"role":"user","content":"Why is the sky blue?"}]) print(r.message.content)★
The main entry point. Returns a typed ChatResponse; the reply is r.message.content.
import ollama ollama.generate(model="llama3.2", prompt="...")
Single-turn completion. Module-level functions use a default client at localhost:11434.
from ollama import Client client = Client(host="http://remote:11434") client.chat(model="llama3.2", messages=[...])★
Use Client to target a non-default host or pass custom headers/timeouts.
from ollama import AsyncClient await AsyncClient().chat(model="llama3.2", messages=[...])
Async twin. Also ollama.list(), pull(), show(), create(), delete(), ps().

8Streaming & optionstoken by token

stream = chat(model="llama3.2", messages=[...], stream=True) for chunk in stream: print(chunk.message.content, end="", flush=True)★
stream=True yields partial ChatResponse chunks — concatenate chunk.message.content.
chat(model="llama3.2", messages=[...], options={"temperature":0.2, "num_ctx":8192, "seed":0})★
Decoding knobs go in options (mirrors the REST options): temperature, top_p, top_k, num_predict, stop, seed.
msg = {"role":"user","content":"describe", "images":["cat.png"]}
Multimodal: attach image paths/bytes via images for vision models (e.g. llama3.2-vision, llava).
r.total_duration · r.eval_count · r.prompt_eval_count
Timing/token stats come back on the response for benchmarking.

9Tools & structured outputfunction calling

def add(a: int, b: int) -> int: """Add two numbers.""" return a + b r = chat("llama3.2", messages=[...], tools=[add])★
Pass Python functions directly as tools; the library builds the schema. Read requested calls from r.message.tool_calls.
for tc in r.message.tool_calls or []: result = fns[tc.function.name](**tc.function.arguments)
You execute the tool, then send the result back as a {"role":"tool"} message for the model to continue.
from pydantic import BaseModel chat("llama3.2", messages=[...], format=Country.model_json_schema())★
Structured output: pass a JSON schema (or format="json") to force valid JSON; parse with Model.model_validate_json(r.message.content).

10OpenAI compatibilitydrop-in

from openai import OpenAI client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")★
Ollama exposes an OpenAI-compatible surface at /v1. Point the OpenAI SDK at it; api_key is required but ignored.
client.chat.completions.create(model="llama3.2", messages=[{"role":"user","content":"Hi"}])★
Works with chat.completions, completions, and embeddings — so any OpenAI-targeting framework runs on Ollama by changing the base URL.
# LangChain: ChatOllama / OllamaEmbeddings (langchain-ollama)
First-class integrations exist too — e.g. from langchain_ollama import ChatOllama.

11Server config & env varstuning

export OLLAMA_HOST=0.0.0.0:11434★
Bind to all interfaces to serve other machines (default is localhost-only). Restart the server after changing env vars.
export OLLAMA_MODELS=/data/models
Move the model store off the system drive (models are large).
export OLLAMA_KEEP_ALIVE=30m★
How long a model stays loaded in memory after use (default 5m). -1 = keep forever; 0 = unload immediately.
export OLLAMA_NUM_PARALLEL=4 · OLLAMA_MAX_LOADED_MODELS=2
Concurrency: parallel requests per model and how many models load at once. OLLAMA_FLASH_ATTENTION=1 to enable flash attention.

12Ecosystemthe wider stack

# GUIs: Ollama desktop app, Open WebUI, Enchanted, LM-friendly IDEs
A native desktop app plus many community front-ends talk to the same :11434 server.
docker run -d -p 11434:11434 ollama/ollama
Official Docker image. Add --gpus all for NVIDIA GPU acceleration inside the container.
# libraries: ollama-python, ollama-js + dozens of community SDKs
Official Python & JavaScript clients; the OpenAI compatibility layer covers the rest.
# works with LangChain, LlamaIndex, Continue, Aider, and more
Because it's local + OpenAI-compatible, Ollama is a common default backend for privacy-sensitive or offline apps.

Ollama cheat sheet