Ollama

tl;dr: ollama run gemma4 and you have a local model. The CLI also launches Claude Code, Codex, OpenCode, Hermes, and OpenClaw against your models, serves a REST API on 11434, and now has a paid cloud tier — local is still free.

What it does

Ollama downloads, quantizes, and serves open models behind one CLI and a local REST API. Backend is llama.cpp (Ollama ships its own fork), with an MLX engine path on Apple Silicon. The CLI banner now advertises Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen, and Gemma as the headline set.

Repo sits at 181k stars, 17.8k forks, MIT. Ollama’s site claims 9M+ developers. Stable line is v0.33.x as of this fetch.

Installation

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

# Windows (PowerShell):
irm https://ollama.com/install.ps1 | iex
# or download OllamaSetup.exe from https://ollama.com/download

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

Running models

# bare `ollama` prompts you to run a model or connect an agent
ollama run gemma4

# launch a coding agent against local models:
ollama launch claude     # Claude Code; also codex, copilot-cli, opencode, deepseek-harness, droid

# turn Ollama into a personal AI assistant across messaging apps:
ollama launch openclaw   # OpenClaw; also hermes, n8n, pi

API

REST API on http://localhost:11434. OpenAI-compatible endpoints plus native /api/chat:

curl http://localhost:11434/api/chat -d '{
  "model": "gemma4",
  "messages": [{"role": "user", "content": "Why is the sky blue?"}],
  "stream": false
}'

Official SDKs: pip install ollama (Python), npm i ollama (JS).

Cloud and pricing

Ollama now sells cloud inference. Pro is $20/mo including $60 of usage; the Max tier includes $300. Models tagged cloud in the library (gpt-oss, glm-5.1, minimax-m2.7, qwen3.5, nemotron-3-super, gemma4 31b) route to Ollama-hosted capacity. Per the privacy page, cloud models are hosted in the US, Europe, and Singapore only; nothing run locally leaves your machine, and prompts are never used for training.

Popular models (pull counts from ollama.com/library)

On this site (named models, not every GGUF): llama3.1, deepseek-r1, qwen3, gemma4, gpt-oss, phi4, deepseek-v3.1, mistral-small3.1, granite3.3, granite3.2-vision, yi, wizardlm2.

Gotchas

• The API binds localhost with no auth by default. Don’t expose 11434 without a proxy.
• cloud-tagged models need an Ollama account and burn plan usage — check the tag before pulling.
• Windows install is install.ps1 via PowerShell, not the curl one-liner.
• Ollama claims 195.6 tok/sec on DeepSeek v4 Flash vs Provider A 97.6 — vendor benchmark, take with salt.
• Cloud usage resets on the 1st of the month; add usage credits if you blow through the included tier.

Related: LM Studio for a GUI agent, llama.cpp for the engine, Open WebUI for a chat UI, Hugging Face Hub for cards and downloads, best local LLM, Ollama vs LM Studio.

Untested here — facts from GitHub and ollama.com fetched 2026-09-10.