Ollama (Local LLM Runtime) Cheat Sheet

Updated 2026-05-21

🧠Study flashcards on this topic110 cards · spaced repetition→

Ollama is an open-source runtime that packages large language models with their configuration and serves them via a local REST API, letting developers run models like Llama, Mistral, Gemma, and Qwen entirely on their own hardware. It solves the privacy, latency, and cost problems of cloud-hosted LLMs by providing a simple CLI, a Docker-friendly server process, and an OpenAI-compatible API that integrates with existing tooling without code changes. The key mental model: Ollama is a model manager and inference server in one — ollama pull downloads, ollama serve exposes port 11434, and every tool that speaks OpenAI's REST dialect can point at http://localhost:11434/v1 and work immediately.

What This Cheat Sheet Covers

This topic spans 16 focused tables and 146 indexed concepts, 110 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core CLI CommandsTable 2: Modelfile InstructionsTable 3: PARAMETER Values (Modelfile & API Options)Table 4: REST API Endpoints (Native)Table 5: OpenAI-Compatible API Endpoints (/v1/*)Table 6: GGUF Quantization LevelsTable 7: GPU Acceleration (CUDA / Metal / ROCm / Vulkan)Table 8: Environment VariablesTable 9: Model Library — Key FamiliesTable 10: Multimodal / Vision CapabilitiesTable 11: Thinking / Reasoning ModeTable 12: Python SDK UsageTable 13: Structured Outputs & Tool CallingTable 14: Embeddings & RAG IntegrationTable 15: Importing Custom Models (GGUF / Safetensors)Table 16: Key Integrations (Open WebUI, Continue, LangChain, LlamaIndex)

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core CLI Commands

Every workflow starts at the command line. These commands cover the full lifecycle of a model — downloading, running, inspecting, and removing it — and are the first things to learn before touching the API or Modelfiles.

Command	Example	Description
ollama pull	`ollama pull llama3.2` `ollama pull llama3.2:3b`	Downloads a model (and specific tag/size) from the Ollama registry into local storage.
ollama run	`ollama run llama3.2` `ollama run gemma3 "Why is the sky blue?"`	Pulls (if needed) then launches an interactive chat session, or runs a one-shot prompt when text is supplied as argument.
ollama list (ollama ls)	`ollama list`	Shows all locally downloaded models with NAME, ID, SIZE, and MODIFIED columns.
ollama show	`ollama show llama3.2`	Displays model metadata: architecture, parameters, template, system prompt, and license.
ollama ps	`ollama ps`	Lists currently loaded models with their VRAM/RAM footprint — useful to diagnose memory pressure.
ollama stop	`ollama stop llama3.2`	Immediately unloads a running model from memory without waiting for the keep-alive timer.
ollama rm	`ollama rm llama3.2`	Permanently removes a model from local storage.
ollama serve	`ollama serve`	• Starts the Ollama API server on port 11434 • use `--help` to view environment variable overrides
ollama create	`ollama create mymodel -f Modelfile`	Builds a new named model from a Modelfile (or from a GGUF/Safetensors path).
ollama cp	`ollama cp llama3.2 alice/llama3.2`	Creates a local alias/copy of a model — necessary before `ollama push` to namespace it under your username.
ollama push	`ollama push alice/llama3.2`	Uploads a namespaced model to ollama.com (requires `ollama signin` first).
ollama launch	`ollama launch claude` `ollama launch opencode --config`	Sets up and starts coding tools (Claude Code, OpenCode, Codex, Droid) with local or cloud models — no manual env-var config needed.
ollama signin / ollama signout	`ollama signin`	Authenticates against ollama.com for push/pull of private models.

Table 2: Modelfile Instructions

A Modelfile is the recipe Ollama uses to create a named model. It controls the base, behavior, personality, and inference parameters baked in at creation time. Most custom workflows only need FROM, SYSTEM, and a few PARAMETER lines.

Instruction	Example	Description
FROM	`FROM llama3.2` `FROM ./model.gguf`	Required. Specifies the base model — either a registry name/tag or a path to a local GGUF or Safetensors directory.

The rest of this section is part of the full cheat sheet

Table 3: PARAMETER Values (Modelfile & API Options)

These parameters tune inference behavior and can be set in a Modelfile (baked in) or passed in the options field of any API request (per-request override). Understanding the most important few — temperature, num_ctx, and top_p — covers 90% of tuning needs.

Parameter	Example	Description
temperature	`PARAMETER temperature 0.7`	• Controls output randomness • 0.0 = deterministic, 1.0 = default creativity, above 1.0 = more chaotic • Range: 0.0–2.0 (default 0.8).

The rest of this section is part of the full cheat sheet

Table 4: REST API Endpoints (Native)

Ollama exposes a REST API on http://localhost:11434. These native endpoints give full control over inference and model management; the OpenAI-compatible /v1/* endpoints (Table 5) are a compatibility layer on top of these. All inference endpoints support streaming JSON by default.

Endpoint	Example	Description
POST /api/chat	`curl http://localhost:11434/api/chat -d '{"model":"llama3.2","messages":[{"role":"user","content":"Hi"}]}'`	• Conversational completion with message history • supports `tools`, `format`, `think`, `stream`, and per-request `options`. Primary endpoint for chat applications.

The rest of this section is part of the full cheat sheet

Table 1: Core CLI Commands

Command	Example	Description
ollama pull	`ollama pull llama3.2` `ollama pull llama3.2:3b`	Downloads a model (and specific tag/size) from the Ollama registry into local storage.
ollama run	`ollama run llama3.2` `ollama run gemma3 "Why is the sky blue?"`	Pulls (if needed) then launches an interactive chat session, or runs a one-shot prompt when text is supplied as argument.
ollama list (ollama ls)	`ollama list`	Shows all locally downloaded models with NAME, ID, SIZE, and MODIFIED columns.
ollama show	`ollama show llama3.2`	Displays model metadata: architecture, parameters, template, system prompt, and license.
ollama ps	`ollama ps`	Lists currently loaded models with their VRAM/RAM footprint — useful to diagnose memory pressure.
ollama stop	`ollama stop llama3.2`	Immediately unloads a running model from memory without waiting for the keep-alive timer.
ollama rm	`ollama rm llama3.2`	Permanently removes a model from local storage.
ollama serve	`ollama serve`	• Starts the Ollama API server on port 11434 • use `--help` to view environment variable overrides
ollama create	`ollama create mymodel -f Modelfile`	Builds a new named model from a Modelfile (or from a GGUF/Safetensors path).
ollama cp	`ollama cp llama3.2 alice/llama3.2`	Creates a local alias/copy of a model — necessary before `ollama push` to namespace it under your username.
ollama push	`ollama push alice/llama3.2`	Uploads a namespaced model to ollama.com (requires `ollama signin` first).
ollama launch	`ollama launch claude` `ollama launch opencode --config`	Sets up and starts coding tools (Claude Code, OpenCode, Codex, Droid) with local or cloud models — no manual env-var config needed.
ollama signin / ollama signout	`ollama signin`	Authenticates against ollama.com for push/pull of private models.

Table 2: Modelfile Instructions

Instruction	Example	Description
FROM	`FROM llama3.2` `FROM ./model.gguf`	Required. Specifies the base model — either a registry name/tag or a path to a local GGUF or Safetensors directory.

Table 3: PARAMETER Values (Modelfile & API Options)

Parameter	Example	Description
temperature	`PARAMETER temperature 0.7`	• Controls output randomness • 0.0 = deterministic, 1.0 = default creativity, above 1.0 = more chaotic • Range: 0.0–2.0 (default 0.8).

Table 4: REST API Endpoints (Native)

Endpoint	Example	Description
POST /api/chat	`curl http://localhost:11434/api/chat -d '{"model":"llama3.2","messages":[{"role":"user","content":"Hi"}]}'`	• Conversational completion with message history • supports `tools`, `format`, `think`, `stream`, and per-request `options`. Primary endpoint for chat applications.