Local & self-hosted appsDevelopment & infrastructureOllamaOpen source
Ollama
Decision matrix
Ollama runtime decision matrix
Overview
Verdict: the default choice when a local model should serve other tools, such as an automation, a chat front end or a script. It installs in minutes and runs quietly in the background. For a chat window on your own laptop, LM Studio or Jan are friendlier; for a team server, add Open WebUI on top.
Other good fits
Running a model on a lab PC or small server that several of your own tools share.
Trying bigger open models through Ollama's paid cloud without changing your setup.
+ Pros
Read pros
Ollama turns your computer into a small model server. One command downloads a model and starts it; from then on any program on the machine can send it requests at http://localhost:11434, in Ollama's own format or the OpenAI format most tools already understand. There is a simple desktop app for chatting too, but Ollama's real strength is being the engine behind other tools. Ollama also sells optional cloud plans for models too large for your hardware.
Pick it when
You are building an automation or app that needs a model, and the data should stay on your machine.
You want one local model shared by several tools: Open WebUI for chat, n8n for workflows, an editor plugin for code.
You like the idea of "local by default, cloud when needed" from the same command line.
− Cons
Read cons
Ollama is deliberately simple. It chooses many settings for you, which is good for a start and limiting when you want fine control over memory and speed. It serves one person or a few tools well; it is not tuned for dozens of simultaneous users. The API listens on 127.0.0.1 only by default. Keep it that way unless you put authentication in front, because the API has no login of its own. Cloud models are convenient, but prompts then go to Ollama's servers (Ollama says it does not log or train on them). Set OLLAMA_NO_CLOUD=1 if everything must stay local.
Pick something else when
You want a chat app with a model browser for yourself: LM Studio or Jan.
You want full control over every model setting: llama.cpp directly.
Many users or a busy workflow need high throughput from one GPU server: vLLM.
Avoid if
You need many users hitting one model at once with high throughput; vLLM is built for that.
Your machine has under 8 GB of free RAM or graphics memory; even the quick-start model will struggle.
Requirements
Install for macOS, Windows or Linux via the quick start (opens in a new tab). Check GPU support (opens in a new tab): Apple Metal, supported NVIDIA/AMD devices and selected Vulkan backends; CPU or mixed placement is possible. For gemma4:e2b, the guide lists a ~7.2 GB download and recommends 8 GB available VRAM/unified memory, not 8 GB total system RAM. Precision, context and concurrency are unspecified in that recommendation; no measured minimum is established here.
Install Ollama (macOS, Windows or Linux) and run ollama run gemma4:e2b, the quick-start model (about 7.2 GB).
Type a test: "Classify this request as ORDER, QUESTION or COMPLAINT. Answer with one word: The centrifuge we bought in May makes a grinding noise."
Collect ten real (anonymised) requests from your inbox and write down the category you would give each.
Send the same ten from a tool: an n8n workflow with an Ollama or OpenAI node pointed at http://localhost:11434, or a short script.
Compare the model's labels with yours.
It worked if at least nine of ten labels match yours and nothing left your computer. If the model misses often, try a larger model before giving up on local.
Integration
The default server binds 127.0.0.1:11434. Use native /api/chat or the supported subset of OpenAI APIs at /v1/chat/completions; local calls need no API key. Cloud models are separate. Set OLLAMA_NO_CLOUD=1 and restart to disable Ollama cloud models/web search. This is not a network sandbox: downloads and other clients may still contact external services. Keep loopback unless you add remote access controls.
Check the result
Expect an answer preserving 12 October and cleaning telescope lenses; compare it with the supplied sentence. ollama ps shows loaded-model CPU/GPU placement if performance is unclear. This is an expected result, not captured output.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Ollama's pricing page, quick start, FAQ and GitHub releases (v0.34.3), checked 23 September 2026. The curriculum team has not installed or run Ollama for this review; the first project is a suggested exercise.
Ollama pricing · Sep 23, 2026 — Free, Pro $20/month ($200/year), Max $100/month, Team $500/month; usage credits; no logging or training on prompts.
Ollama quick start · Sep 23, 2026 — ollama run gemma4:e2b, about 7.2 GB download, 8 GB available VRAM recommended; local API on port 11434.
Ollama FAQ · Sep 23, 2026 — Binds 127.0.0.1:11434 by default; OLLAMA_NO_CLOUD=1 local-only mode; local prompts not seen by Ollama.
Ollama releases · Sep 23, 2026 — Current release v0.34.3.
Further review notes & alternatives
Costs and operating effort
Prices checked 23 September 2026:
Plan
Price
Cloud usage included
Free
$0
Run models locally; starter cloud credits, 1 concurrent cloud request
Pro
$20/month or $200/year
$60 of cloud credits per month, 3 concurrent requests
Max
$100/month
$300 of cloud credits per month, 10 concurrent requests
Team
$500/month (early access)
$1,000 of shared credits, unlimited users
Local use is free, open-source software (MIT); you pay for the machine. The quick start recommends 8 GB of free graphics or unified memory for its small default model; bigger models need proportionally more, so larger ones call for a desktop GPU or a Mac with plenty of unified memory. Upkeep: update Ollama now and then, pull newer models, and delete old ones from disk.
