cyberivy
OllamaLocal AILocal LLMOpen Source AIDeveloper ToolsSelf-hosted AIPrivacyClaude Code

Ollama runs open language models on your own computer, and in the cloud

November 19, 2026

Schwarz-weiße Strichzeichnung eines freundlich blickenden Lamakopfs mit aufgestellten Ohren auf weißem Hintergrund

Ollama launches open AI models locally with a single command and offers compatible APIs. What the MIT-licensed tool can do, what the cloud plans cost and where the limits are.

What this is about

Ollama is one of the best-known tools for running open language models on your own computer. The project is MIT licensed, and the GitHub repository ollama/ollama had around 187,000 stars on 19 November 2026. The current version, v0.40.16, was released on 9 November 2026. Ollama is no longer just a command-line tool for tinkerers: there are now desktop apps for macOS, Windows and Linux, compatible interfaces for OpenAI and Anthropic clients, and a paid cloud offering for models that will not run on ordinary hardware.

That mix of local and cloud is exactly why a sober look makes sense: what do you get for free, what do you pay for, and where are the limits?

What Ollama actually does

At its core, Ollama downloads a model from its library, manages it locally and serves it through a small server on the machine. A command such as ollama run gemma4 is enough to chat with a model in the terminal. Under the hood Ollama uses, among other things, llama.cpp and supports Nvidia and AMD GPUs as well as Apple Silicon.

According to the documentation, Ollama can be called through its own API as well as through OpenAI- and Anthropic-compatible endpoints. This lets many existing programs that were built for a cloud service point at a local model without major changes. There are official libraries for Python and JavaScript.

A newer addition is the ollama launch command: it starts coding agents such as Claude Code and connects them to a selected Ollama model. According to the changelog, release v0.40.16 of 9 November 2026 makes sure that ollama launch claude always routes Claude Code to the chosen model, and explains clearly when a cloud model has been retired.

Why it matters

For many teams the key question is not "Which model is best?" but "Is this data allowed to leave the machine at all?". Ollama gives a simple answer: local models run entirely on your own hardware. That is relevant for developers, research groups, law firms or medical practices that want to build prototypes without sending confidential text to a provider.

At the same time, Ollama has built a cloud offering. According to its pricing page there is a pay-as-you-go plan with no base fee, where you pay per token for cloud models, a Pro plan at 20 US dollars a month with 60 US dollars of usage credits, and a Team plan at 500 US dollars a month with 1,000 US dollars of shared credits. Outside 12:00 to 18:00 UTC on weekdays and all day at weekends, cheaper off-peak prices apply to some models. Anyone who starts locally and later needs more compute does not have to switch tools.

Ollama is also a building block in many other tools. Interfaces such as Open WebUI or research tools such as Khoj, both previously covered by Cyber Ivy, can use Ollama as their model backend.

In plain language

Ollama is like a capsule coffee machine for your kitchen at home. You insert a capsule, press a button and get coffee without grinding beans or setting water temperatures. The capsules are the models; the machine takes care of the rest. And when guests arrive and your own machine is not enough, you can order coffee from the same manufacturer, but you have to pay for it and it leaves the house.

A practical example

A tax advisory office with eight employees wants to test whether a language model can summarise client letters. The data must not leave the office. One employee installs Ollama on a workstation with 32 GB of RAM and downloads a mid-sized open model. Using a lightweight interface, she summarises ten anonymised sample letters and compares the results with her own notes.

For seven out of ten letters the summaries are usable; for three, important deadlines are missing. The office decides: the model may produce drafts, but every deadline is checked by a person. Up to this point the test has cost only time and electricity, no subscription.

Scope and limits

  • Hardware determines quality: local models are limited by memory and graphics card. Small models run on laptops but often deliver weaker results on complex tasks than large cloud models.
  • Cloud means sharing data again: as soon as you use cloud models through Ollama, your data leaves your own machine. Anyone using Ollama for privacy reasons needs to keep a clear line between which model runs locally and which runs in the cloud.
  • The local server needs protection: the Ollama server is designed for your own machine and has no built-in user management. Anyone who makes it reachable on a network or the internet should put it behind a reverse proxy with authentication or a firewall. Security researchers repeatedly found publicly reachable AI servers in 2026.

Ollama is also not a finished end-user assistant with document management or a permissions model. It is more the engine around which such applications are built.

SEO & GEO keywords

Ollama, local AI, local LLM, open-source language models, llama.cpp, Ollama Cloud, OpenAI-compatible API, Claude Code, self-hosted AI, privacy, Gemma, gpt-oss

💡 In plain English

Ollama is a free program that lets you use AI language models on your own computer without your text leaving the device. If you need bigger models, you can pay to use them in Ollama's cloud. In that case the usual privacy questions of any cloud service apply.

Key Takeaways

  • →Ollama is MIT licensed and runs on macOS, Windows and Linux.
  • →Version v0.40.16 was released on 9 November 2026 and improves, among other things, the ollama launch command for Claude Code.
  • →Ollama offers its own API as well as OpenAI- and Anthropic-compatible endpoints.
  • →Local use is free; cloud models are billed per token, and Pro costs 20 US dollars a month.
  • →Using cloud models means sharing data with an external service.
  • →The local server should not be exposed to the network without protection.

FAQ

Is Ollama free?

The software and local use are free and MIT licensed. Costs only arise for cloud models, either per token or through the Pro and Team plans.

What hardware do I need?

Small models run on current laptops. For larger models, plenty of memory and a capable graphics card or Apple Silicon help.

Does my data stay local?

With local models, yes. If you choose a cloud model, requests are sent to Ollama's servers.

Can I connect existing programs?

Yes, Ollama offers OpenAI- and Anthropic-compatible interfaces as well as libraries for Python and JavaScript.

Sources & Context