Ollama

Install and Setup

macOS:

brew install ollama

Linux:

curl -fsSL https://ollama.com/install.sh | sh

Windows: Download from https://ollama.com/download/windows.

Start the server:

ollama serve

The server listens on http://localhost:11434 by default. Set OLLAMA_HOST to change the bind address.

Pull and Run Models

ollama pull llama3       # download without running
ollama run llama3        # run (auto-pulls if missing)
ollama list              # list downloaded models

Interactive Chat vs One-Shot Generation

Interactive chat (opens a REPL, type /bye to exit):

ollama run llama3

One-shot generation (pipe input, get output, exit):

echo "Explain quicksort in two sentences" | ollama run llama3
cat main.py | ollama run codellama "Review this code for bugs"

REST API

Generate (completion)

curl http://localhost:11434/api/generate -d '{
  "model": "llama3",
  "prompt": "Why is the sky blue?",
  "stream": false
}'

Chat (multi-turn)

curl http://localhost:11434/api/chat -d '{
  "model": "llama3",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is 2+2?"}
  ],
  "stream": false
}'

Embeddings

curl http://localhost:11434/api/embed -d '{
  "model": "nomic-embed-text",
  "input": "Ollama is a tool for running local LLMs"
}'

Set "stream": true (the default) to receive newline-delimited JSON chunks.

Model Management

ollama list                    # list downloaded models
ollama show llama3             # show model details (parameters, template, license)
ollama cp llama3 my-llama3     # copy/alias a model
ollama rm my-llama3            # delete a model
ollama ps                      # list currently loaded/running models

ollama ps shows VRAM usage, quantization level, and time until unload.

Modelfile

A Modelfile defines a custom model:

FROM llama3

SYSTEM "You are a senior software engineer. Be concise. Provide code examples."

PARAMETER temperature 0.3
PARAMETER num_ctx 8192
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
PARAMETER stop "<|eot_id|>"

Key parameters:

• temperature -- randomness (0.0 = deterministic, 1.0+ = creative).
• num_ctx -- context window in tokens. Higher values use more VRAM.
• top_p -- nucleus sampling threshold.
• top_k -- limits token selection pool.
• repeat_penalty -- penalizes repeated tokens.
• stop -- stop sequence(s).
• num_gpu -- layers to offload to GPU (0 for CPU-only).

Create Custom Models from Modelfile

ollama create my-coder -f ./Modelfile
ollama run my-coder

To update, edit the Modelfile and run ollama create again with the same name.

GPU vs CPU Detection and Configuration

Ollama auto-detects NVIDIA (CUDA), AMD (ROCm), and Apple Silicon (Metal) GPUs.

ollama ps   # PROCESSOR column shows gpu or cpu

Force CPU-only:

CUDA_VISIBLE_DEVICES="" ollama serve    # per-session
OLLAMA_NUM_GPU=0 ollama serve           # server-wide

Per-model GPU control in a Modelfile:

PARAMETER num_gpu 0    # force CPU
PARAMETER num_gpu 999  # offload all layers to GPU (default)

For multi-GPU, set CUDA_VISIBLE_DEVICES=0,1.

Popular Models and When to Use Which

Model	Size	Best for
llama3 (8B)	4.7 GB	General chat, reasoning, instruction following
llama3:70b	40 GB	Higher quality when you have the VRAM
codellama (7B)	3.8 GB	Code generation, completion, infilling
mistral (7B)	4.1 GB	Fast general-purpose, structured output
phi3 (3.8B)	2.2 GB	Small footprint, good quality for its size
gemma2 (9B)	5.4 GB	Strong reasoning, multilingual
deepseek-coder-v2 (16B)	8.9 GB	Code generation, multi-language
nomic-embed-text	274 MB	Text embeddings for RAG
llava (7B)	4.7 GB	Multi-modal image understanding

For constrained hardware (8 GB RAM), use phi3 or llama3 with q4 quantization.

Embeddings for RAG

Use nomic-embed-text (768 dimensions) or mxbai-embed-large (1024 dimensions).

curl http://localhost:11434/api/embed -d '{
  "model": "nomic-embed-text",
  "input": ["document chunk one", "document chunk two"]
}'

Response contains an "embeddings" array of float vectors. Store in a vector database (ChromaDB, pgvector, Qdrant, FAISS) for similarity search.

Integration with Python

Using requests

import requests

response = requests.post("http://localhost:11434/api/generate", json={
    "model": "llama3",
    "prompt": "Explain monads simply",
    "stream": False
})
print(response.json()["response"])

Using ollama-python (pip install ollama)

import ollama

# Chat
response = ollama.chat(model="llama3", messages=[
    {"role": "user", "content": "Explain monads simply"}
])
print(response["message"]["content"])

# Embeddings
result = ollama.embed(model="nomic-embed-text", input="some text")
print(len(result["embeddings"][0]))  # 768

# Streaming
for chunk in ollama.chat(model="llama3", messages=[
    {"role": "user", "content": "Write a haiku"}
], stream=True):
    print(chunk["message"]["content"], end="", flush=True)

Integration with JavaScript

Using fetch

const response = await fetch("http://localhost:11434/api/generate", {
  method: "POST",
  body: JSON.stringify({ model: "llama3", prompt: "Explain closures", stream: false }),
});
const data = await response.json();
console.log(data.response);

Using ollama-js (npm install ollama)

import { Ollama } from "ollama";
const ollama = new Ollama({ host: "http://localhost:11434" });

const response = await ollama.chat({
  model: "llama3",
  messages: [{ role: "user", content: "Explain closures" }],
});
console.log(response.message.content);

Integration with LangChain

Python (pip install langchain-ollama):

from langchain_ollama import ChatOllama, OllamaEmbeddings

llm = ChatOllama(model="llama3", temperature=0.3)
response = llm.invoke("What is dependency injection?")

embeddings = OllamaEmbeddings(model="nomic-embed-text")
vectors = embeddings.embed_documents(["first chunk", "second chunk"])

JavaScript (npm install @langchain/ollama):

import { ChatOllama } from "@langchain/ollama";
const llm = new ChatOllama({ model: "llama3", temperature: 0.3 });
const response = await llm.invoke("What is dependency injection?");

Multi-Modal Models

Use llava or llava-llama3 for image understanding:

curl http://localhost:11434/api/chat -d '{
  "model": "llava",
  "messages": [{"role": "user", "content": "Describe this image", "images": ["BASE64_DATA"]}],
  "stream": false
}'

import ollama, base64

with open("photo.jpg", "rb") as f:
    img = base64.b64encode(f.read()).decode()

response = ollama.chat(model="llava", messages=[
    {"role": "user", "content": "What do you see?", "images": [img]}
])

Performance Tuning

Context size -- use the smallest that fits your workload:

PARAMETER num_ctx 4096   # default
PARAMETER num_ctx 32768  # long documents, more VRAM

Partial GPU offloading when model exceeds VRAM:

PARAMETER num_gpu 20     # 20 layers on GPU, rest on CPU

Batch size for prompt processing speed:

PARAMETER num_batch 512  # default, increase for faster eval

Keep-alive control (how long model stays loaded):

curl http://localhost:11434/api/generate -d '{
  "model": "llama3", "prompt": "hi", "keep_alive": "30m"
}'

Use "keep_alive": 0 to unload immediately, -1 to keep indefinitely.

Server-wide environment variables:

• OLLAMA_MAX_LOADED_MODELS -- concurrent models in memory (default 1).
• OLLAMA_NUM_PARALLEL -- concurrent requests per model.
• OLLAMA_FLASH_ATTENTION=1 -- reduce memory with flash attention.

Running as a Service

macOS (Homebrew):

brew services start ollama

Linux (systemd -- created automatically by the install script):

sudo systemctl enable --now ollama
sudo journalctl -u ollama -f

Customize with sudo systemctl edit ollama:

[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
Environment="OLLAMA_NUM_PARALLEL=4"

Then sudo systemctl daemon-reload && sudo systemctl restart ollama.

Docker:

docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
docker exec -it ollama ollama run llama3

With NVIDIA GPU: add --gpus=all to the run command.