Complete Ollama command reference: install, pull models, chat, OpenAI-compatible API server, Modelfile and environment variables.

💡 What You Will Learn

Complete Ollama command reference: install, pull models, chat, OpenAI-compatible API server, Modelfile and environment variables.

📜 Table of Contents

• Install
• Model Management
• Chat
• API Server
• Modelfile (custom models)
• Environment Variables
• Key Parameters
• FAQ

Run Ollama Locally: Every Command You Need

Ollama (177k+ stars on GitHub) is the easiest way to run local LLMs in 2026. This is a copy-paste command reference.

Install

• macOS/Linux: curl -fsSL https://ollama.com/install.sh | sh
• Windows: download the installer from ollama.com
• Docker: docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama

Verify with ollama --version.

Model Management

Models are named name:tag (e.g. llama3.2:3b, qwen2.5:7b).

Chat

ollama run qwen2.5:7b

In-chat shortcuts: /bye to exit, /help for all shortcuts, /set parameter num_ctx 8192 to raise context. One-shot: ollama run qwen2.5:7b "explain RAG in one sentence".

API Server

ollama serve listens on port 11434 and exposes an OpenAI-compatible /v1/chat/completions endpoint, so any OpenAI SDK can point at http://localhost:11434/v1. Native endpoints: /api/generate (single completion) and /api/chat (multi-turn). ollama ps shows models loaded in memory.

Modelfile (custom models)

FROM qwen2.5:7b
SYSTEM "You are a careful technical editor."

Save as Modelfile, run ollama create my-editor -f Modelfile, then ollama run my-editor. Useful directives: PARAMETER temperature 0.3, PARAMETER num_ctx 8192.

Environment Variables

• OLLAMA_KEEP_ALIVE: how long a model stays in memory (e.g. 30m, -1 = always)
• OLLAMA_MODELS: model storage directory
• OLLAMA_HOST: bind address (0.0.0.0:11434 for LAN access)

Key Parameters

num_ctx (context window, raise for long documents), temperature (0.8 default), top_p (0.9), num_predict (max output tokens, -1 unlimited).

FAQ

Q: Can I run without a GPU? A: Yes, CPU runs 7B at about 2-5 token/s; GPU is typically 5-20x faster. Q: Long documents get truncated? A: Raise num_ctx (--num-ctx 8192 or in Modelfile); larger context uses more memory. Q: LAN access fails? A: Set OLLAMA_HOST=0.0.0.0:11434 and open the firewall port.

Note: model tags and versions change over time; check the official docs.

❓ FAQ

Does Ollama need a GPU?

No - it runs on CPU, but GPU (CUDA/Metal) gives 5-20x speedup.

Is the API OpenAI-compatible?

Yes, the /v1/chat/completions endpoint works with most tools that expect OpenAI.