Local LLMs with Ollama: Installation and Usage

Ollama is a tool for running large language models (LLMs) locally on your computer or server. It downloads models, manages them, and provides a CLI and HTTP API without a cloud subscription or sending data to third parties. Below - why local models matter, how to install Ollama on Windows, macOS, and Linux, which models to choose, and how to connect them to your apps.

• Privacy - prompts and documents never leave your machine
• Offline - works without internet after the model is downloaded
• Free - no per-token fees, only your hardware and electricity
• API - OpenAI-compatible format for Cursor, n8n, LangChain, and other tools

What is Ollama

Ollama is an open-source runtime for local LLMs. It simplifies what used to require manual setup of llama.cpp, CUDA, Python environments, and weight conversion: one command ollama pull llama3 downloads a model, ollama run llama3 starts an interactive chat.

Under the hood, Ollama uses llama.cpp and its own Modelfile format - like a Dockerfile for models. You can:

• run ready-made models from the Ollama library;
• create custom variants with system prompts, parameters, and adapters;
• call the model via REST API at http://localhost:11434.

Ollama does not replace cloud flagships like GPT-5.6 or Claude Fable 5 for complex agent tasks, but covers a large set of scenarios: draft writing, summarization, classification, RAG over internal documents, prompt prototyping, and development without burning API credits.

Why run LLMs locally

Criterion	Cloud API	Ollama (local)
Privacy	Data on provider servers	Data stays on your machine
Cost	Pay per token	Free (except hardware)
Latency	Depends on network	Minimal on localhost
Top model quality	Higher	Lower for open-weight models
Offline	No	Yes, after weights are downloaded

Local models are especially useful for:

• Corporate data - contracts, email, codebases you cannot send to SaaS.
• Development - testing RAG and agent pipelines without an API bill.
• Edge and air-gapped - servers with no internet, labs, field deployments.
• Experiments - quick model and parameter swaps without changing providers.

Hardware requirements

RAM and GPU availability determine which model you can run comfortably.

Model (example)	Parameters	RAM (CPU)	VRAM (GPU)
Llama 3.2	3B	4-8 GB	4 GB
Qwen 3	8B	8-16 GB	6-8 GB
Llama 3.1	8B	8-16 GB	6-8 GB
Mistral	7B	8-16 GB	6-8 GB
Qwen 3	32B	32+ GB	20-24 GB
Llama 3.1	70B	64+ GB	40-48 GB

CPU-only works but is slow: a 7B model on 16 GB RAM without a GPU yields a few tokens per second. A GPU (NVIDIA with CUDA, Apple Silicon with Metal) speeds up generation dramatically. On Mac with M1/M2/M3/M4, Ollama uses Metal out of the box; on Windows and Linux you need an NVIDIA GPU with an up-to-date driver.

If RAM is tight - pick quantized models (Q4, Q5): they use less memory with a small quality trade-off.

Installation

Windows

1. Download the installer from ollama.com/download.
2. Run the .exe and finish the setup wizard.
3. Ollama starts as a background service; an icon appears in the system tray.
4. Open PowerShell or cmd and verify:

ollama --version

macOS

brew install ollama

Or download the .dmg from the site. On Apple Silicon, Metal is enabled automatically.

Linux

Official script:

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

For a systemd service after install:

sudo systemctl enable ollama
sudo systemctl start ollama

Docker

If you need isolation or server deployment:

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

For GPU on Linux, add --gpus all.

First steps: CLI

Download and run a model

ollama pull llama3.2
ollama run llama3.2

pull downloads weights to ~/.ollama/models (or %USERPROFILE%\.ollama\models on Windows). run opens an interactive chat. Exit with /bye or Ctrl+D.

Useful commands

ollama list              # installed models
ollama ps                # currently running
ollama rm llama3.2       # remove a model
ollama show llama3.2     # parameters and Modelfile
ollama pull qwen3:8b     # specific tag

Tags define size and quantization: llama3.1:8b, mistral:7b-instruct-q4_K_M, qwen3:32b. See the model page in the Ollama library for available tags.

One-shot request without chat

ollama run llama3.2 "Explain what RAG is in three sentences"

HTTP API

Ollama serves on port 11434. The format is close to OpenAI Chat Completions - many clients connect without changes.

Chat (streaming)

curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages": [{"role": "user", "content": "Hello!"}],
  "stream": true
}'

Generate (simple prompt)

curl http://localhost:11434/api/generate -d '{
  "model": "llama3.2",
  "prompt": "Write a haiku about local models",
  "stream": false
}'

OpenAI-compatible endpoint

With versions supporting /v1/chat/completions:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama3.2",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

In Cursor, Continue, Open WebUI, and n8n set Base URL to http://localhost:11434/v1 and the model name from ollama list.

Which models to choose

Guidelines for 2026:

Task	Models	Notes
Fast chat, weak hardware	Llama 3.2 3B, Qwen 3 4B, Phi-4 mini	Minimum RAM
General balance	Llama 3.1 8B, Mistral 7B, Qwen 3 8B	Sweet spot for 16 GB
Code and instructions	Qwen 3 Coder, DeepSeek Coder V2, Codestral	Better at SQL, Python, diffs
Multilingual	Qwen 3, Llama 3.1	Qwen is strong multilingual
Long context	Qwen 3, Llama 3.1 8B (128K+)	RAG and large documents
Max local quality	Qwen 3 32B, Llama 3.1 70B, Mistral Large	Needs powerful GPU or lots of RAM

Start with Qwen 3 8B or Llama 3.1 8B - acceptable quality on a typical 16 GB laptop. For production RAG, compare 2-3 models on your real queries: site benchmarks do not replace your domain.

Modelfile: custom model

A Modelfile sets system prompt, temperature, context, and base model:

FROM llama3.2

PARAMETER temperature 0.3
PARAMETER num_ctx 8192

SYSTEM """
You are a technical support assistant.
Answer briefly, step by step, in English.
If you do not know the answer - say so.
"""

Create and run:

ollama create support-bot -f Modelfile
ollama run support-bot

This fixes behavior for the team without changing client code.

Integrations

Cursor and IDEs

In model settings, set OpenAI-compatible endpoint: http://localhost:11434/v1, API key can be empty or ollama. Model - name from ollama list. Local models suit drafts and refactoring; complex agentic coding still favors cloud Codex or Claude Code.

Open WebUI

Open WebUI - a ChatGPT-style web UI on top of Ollama. Good for teams: chat history, PDF upload, RAG built in.

docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data --name open-webui \
  ghcr.io/open-webui/open-webui:main

n8n, LangChain, LlamaIndex

• n8n - Ollama Chat Model node or HTTP Request to /api/chat.
• LangChain - ChatOllama(base_url="http://localhost:11434", model="qwen3:8b").
• LlamaIndex - Ollama as LLM and embedding (if the model supports it).

Local RAG

Typical stack: Ollama (LLM + embeddings via nomic-embed-text or mxbai-embed-large) + ChromaDB or Qdrant + your documents. Everything works offline after indexing.

Operations tips

1. Keep one model in memory - ollama ps shows loaded models; ollama stop <name> frees RAM.
2. num_ctx - increase context only when needed; larger context = more memory and slower inference.
3. Update Ollama - new releases bring speedups and model support.
4. Monitoring - on servers watch GPU temperature and swap; low RAM causes thrashing.
5. Do not confuse privacy with security - a local model does not send data outward, but malicious prompts or a vulnerable agent are still risky on your machine.

Limitations

• Open-weight 8B models trail top cloud flagships on hard reasoning and long agent chains.
• Large models (32B+) need expensive hardware; cloud may be cheaper than a $2000+ GPU if load is sporadic.
• Multimodal (images, audio) in Ollama is model-dependent - check tag support.
• Weight updates are your responsibility; cloud providers patch models for you.

For many tasks, Ollama + 8B model + RAG is a practical balance of quality, cost, and privacy.

Summary

Ollama lowers the barrier to local LLMs to a few commands: install, pull, run. The API is OpenAI-ecosystem compatible, so the same model works in CLI, web UI, IDE, and automations. Start with Qwen 3 8B or Llama 3.1 8B on your hardware, tune a Modelfile for your scenario, and add RAG if you need facts from your documents - without sending data to the cloud.

Frequently Asked Questions

Do I need a GPU for Ollama?

No, Ollama runs on CPU-only, but inference will be much slower. For comfortable chat with 7B-8B models, a GPU (NVIDIA CUDA) or Apple Silicon with Metal is recommended. On 16 GB RAM without GPU, choose light 3B-4B models or aggressive Q4 quantization.

Where are downloaded models stored and how large are they?

By default in ~/.ollama/models (Linux/macOS) or %USERPROFILE%\.ollama\models (Windows). Size depends on the model: 7B-8B in Q4 is roughly 4-5 GB, 70B is tens of GB. Remove with ollama rm <name>. Override path with OLLAMA_MODELS.

Can I use Ollama in Docker on Windows with GPU?

Yes, via WSL2 and NVIDIA Container Toolkit: Docker Desktop with WSL2 backend, NVIDIA driver for WSL, ollama/ollama image with --gpus all. Native Docker Desktop without WSL2 usually does not expose GPU to Ollama. Simpler alternative - native Ollama install for Windows from the official site.

How do I connect Ollama to Cursor or another OpenAI API client?

Set Base URL to http://localhost:11434/v1, API key to any string (e.g. ollama), model to the exact name from ollama list (e.g. qwen3:8b). Ensure Ollama is running and the model was pulled. If the client runs in Docker, use host.docker.internal:11434 instead of localhost.

Is it safe to send confidential data to a local model?

Data does not go to Ollama Inc. servers during local inference - processing happens on your machine. But app logs, Open WebUI chat history, and disk backups may store prompts. For strict compliance, set retention policies, encrypt disks, and disable third-party UI telemetry. Local inference does not remove leak risks from VM snapshots and shared server accounts.