Ollama
Ollama runs open-weight language models directly on your own machine through a free MIT-licensed engine, and offloads larger ones to its paid cloud. Available for macOS, Windows and Linux through a command line, a desktop app and an API.
What is Ollama?
Ollama is two things under one name, and telling them apart matters. The first is a free, MIT-licensed engine you download and run on your own computer: it fetches open-weight models and serves them locally, with no API key, no rented server and no metering. Its public repository carries 178,300 stars and 17,400 forks. It installs on macOS 14 Sonoma or later, Windows and Linux, and ships as an official Docker image. Three surfaces drive it: a command line, a desktop application, and an HTTP API on port 11434.
The second is Ollama Cloud, a paid hosted service for models too large for a personal machine. Cloud models behave exactly like local ones, pulled and started with the same commands, with the computation offloaded transparently. They can also be reached directly on ollama.com with an API key, through official Python and JavaScript libraries, and the company advertises compatibility with Anthropic's Messages API and with OpenAI's Codex CLI.
Between the two sits a public model library covering GLM, DeepSeek, Kimi, Nemotron, MiniMax, Gemma and Qwen among others, filterable by cloud, embedding, vision, tools and thinking, with download counters running into the tens of millions. Documented capabilities include streaming, reasoning, structured outputs, vision, embeddings, tool calling and web search. Twenty-four integrations are listed, from terminal agents such as Claude Code, OpenCode and Codex to editors including VS Code, JetBrains, Xcode and Zed.
The pitch is ownership and privacy. In local mode nothing leaves the machine, and a documented switch disables cloud features altogether for offline work. Ollama states that prompts and responses are never logged and never used to train models, whichever mode you choose. Model weights and the engine's source code are both open.
The publisher, Ollama Inc., says it serves 8.9 million developers and that 85% of the Fortune 500 use the tool. Local use is free and unlimited; the cloud is metered, starting at 20 US dollars a month.
What it does
- Run open-weight models locally in a single command
- Offload models too large for your hardware to Ollama's cloud
- Call models through a local or a remote HTTP API
- Set up and launch a coding agent straight from the terminal
- Work entirely offline with cloud features switched off
- Publish, share and import your own models
- Use tool calling, vision, embeddings and structured outputs
When to use Ollama / When not to
A quick filter to help you decide if Ollama is the right fit.
When to use Ollama
- Developers who want to run open models without an API key, a server to rent or a per-token bill
- Teams under confidentiality constraints, where prompts must never leave the machine
- Engineers working in disconnected or air-gapped environments
- Users of coding agents such as Claude Code, OpenCode, Codex, Cline or Zed who want to point them at open models
- High-volume workloads where metered pricing would spiral, since local inference is unlimited
When not to use Ollama
- Anyone expecting a chat window in the browser: there is none, and ollama.com is only an account portal
- Mobile-first users, since no iOS or Android application exists
- Machines without a capable GPU, at least for the larger local models
- Buyers who require a signed DPA, a subprocessor list or a SOC 2 or ISO 27001 certificate, none of which are published
- Anyone under 18, whom the terms of service exclude outright
How to use Ollama
A typical end-to-end flow, from setup to results.
- Install from the download page, or paste the one-line install script into a terminal
- Pick a model in the public library, filtering by cloud, vision, tools or embeddings
- Fetch it with ollama pull, then start it with ollama run
- Create an account and run ollama signin to unlock cloud models
- Pull and run a cloud model exactly as you would a local one
- For direct API access, generate a key in your account settings and export it as OLLAMA_API_KEY
- Call the local endpoint on port 11434, or the remote one on ollama.com, with curl or the Python and JavaScript libraries
- Run ollama launch to set up a coding agent such as Claude Code, OpenCode or Codex without editing config files
- Tune the server through environment variables: listening address, CORS origins, proxy, model storage location
- Switch the cloud off entirely with OLLAMA_NO_CLOUD if you need to work offline
Pros & Cons
Pros
- Free and open source under the MIT licence, with unlimited local use and no token bill
- A no-training commitment stated three times over, in the privacy policy, the terms and the pricing FAQ
- One interface for local and cloud, so a workload moves between them without a rewrite
- Genuine, documented offline operation through a single configuration switch
- A large integration ecosystem and compatibility with widely used third-party APIs
- Documentation that covers operations, not just the first five minutes
- Visibly active: blog posts and model updates dated the day before this review
Cons
- No postal address and no legal notice page anywhere on the site
- A single email address for support, legal, GDPR and copyright matters alike
- No DPA, no subprocessor list, no Article 27 EU representative and no published certification
- No browser chat interface and no mobile application
- Cloud usage limits are never quantified in tokens, which makes budgeting guesswork
- New Max sign-ups are suspended for lack of capacity, and the Team plan is still waitlisted
- Extra usage credits expire one year after they are added to the account
Pricing & Plans
A permanent free plan is available at no cost, and local use under it is unlimited. The cheapest paid tier is Pro, at USD 20 per month, or USD 200 per year when billed annually. Paid tiers govern cloud usage only: the downloadable engine itself remains free of charge.
- Free — USD 0
- Pro — USD 20 per month
- or USD 200 per year billed annually
- Max — USD 100 per month
- new sign-ups currently suspended
- Team — USD 25 per seat per month
- five-seat minimum
- waitlist
- Enterprise — custom pricing on request
Data, GDPR & hosting
A consolidated view of how Ollama handles your data.
GDPR overview
Ollama never claims GDPR compliance in so many words, which is why this record marks it as not compliant, but its privacy policy is far from silent on the subject. It names GDPR legal bases, namely contract performance, legitimate interests and consent; it sets out the rights of EU residents, including the right to complain to a supervisory authority; and it commits to notifying that authority within 72 hours of a breach. Access, rectification, erasure, objection, portability and withdrawal of consent are all listed, and an account can be deleted from the profile settings at any time. Transfers to the United States are acknowledged under unnamed appropriate safeguards. What is missing is the machinery: no Article 27 EU representative, no data protection officer, no data processing agreement and no subprocessor list. Rights are exercised through a single address. The policy is dated March 2026.
Who owns the data?
Ollama Inc. is the data controller. Section 7 of the terms leaves ownership of everything you create with you: the company takes only a limited, non-exclusive licence to process that content for the sole purpose of delivering the service, and states plainly that it claims no ownership over your inputs or outputs. What Ollama does keep is its own side of the bargain, namely rights over its software, service and trademarks. Run models locally and the question barely arises, because the privacy policy says the company does not collect, store, transmit or have access to prompts, responses or interactions processed on your device.
Reuse rights
You may reuse your own inputs and outputs freely and without asking permission, subject only to the lawful-use restrictions in section 4 of the terms. On Ollama's side the permitted use is deliberately narrow. Inputs and outputs are never used to train AI models, a commitment repeated in the privacy policy, in the terms and in the pricing FAQ. Cloud prompts and responses are processed transiently and are not logged. What the company does collect is account details, payment data handled by Stripe, support exchanges, and device and usage metadata that explicitly excludes prompt and response content. Data is not sold, and it is not shared for cross-context behavioural advertising. A documented local-only mode, set through disable_ollama_cloud or the OLLAMA_NO_CLOUD variable, stops anything leaving the machine at all, at the cost of cloud models and web search.
Data retention & training
Hosting summary
Where your data lives depends entirely on which half of the product you use. Run a model locally and nothing is hosted at all: the privacy policy states that prompts, responses and model interactions never leave your device. Use a cloud model and processing moves to Ollama's infrastructure, which the company describes as primarily in the United States, with traffic routed to Europe and Singapore for extra capacity. Ollama does not own that infrastructure. Models are hosted with partner NVIDIA Cloud Providers, on whom the company says it imposes no-logging, no-training and zero-retention terms. Separately, the privacy policy acknowledges that personal data may be transferred to and processed in the United States, invoking appropriate safeguards for international transfers without naming them. The website itself resolves to an anycast address on Google's network, geolocated in the United States. No data residency option, hosting certification or subprocessor list is published.
Things to keep in mind
Risks and trade-offs to weigh before adopting Ollama.
- Two very different privacy regimes share one name: local inference sends nothing anywhere, while a cloud model sends your prompt to a third party. It is easy to forget which one you have just switched to.
- The company publishes no postal address and no legal notice, so Ollama Inc. is identifiable by name alone, which is thin if you ever need to enforce anything.
- A single mailbox handles support, legal, GDPR and copyright requests, a narrow channel when something goes wrong.
- No DPA, no subprocessor list and no Article 27 EU representative: European buyers will have to request all three before any serious deployment.
- The privacy policy contradicts itself on the minimum age, stating 18 and then 13; only the terms of service settle the question, at 18.
- Running capable models on your own machine invites unchecked trust in their answers. The terms themselves warn that AI output may be inaccurate or incomplete and should be verified independently before any consequential decision.
- Disputes are pushed to individual arbitration in San Francisco under Californian law, with class actions expressly waived.
Setup & Integrations
Technical difficulty
Easy to install, developer-oriented to use. Installation is one command or a downloaded installer on macOS, Windows and Linux, with automatic updates on macOS and Windows. The entry point is then a terminal, since ollama run starts a model and there is no browser chat interface to fall back on. Large local models need a capable GPU and ample memory, which is precisely why the cloud tier exists. Anything beyond first use, such as proxies, CORS origins, Docker GPU passthrough or changing the listening address, goes through environment variables that differ from one operating system to the next.
Deployment
Integrations
Behind Ollama
Fundraising
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Is Ollama free?
Are my prompts used to train AI models?
Can I use Ollama offline?
Where are the cloud models hosted?
Is there a mobile app or a chat interface in the browser?
Does Ollama provide a DPA or a list of subprocessors?
What is the minimum age?
What are the cloud usage limits?
Should you pick Ollama?
Ollama is a real, widely adopted tool with an identified and well-funded publisher behind it, and its value comes from the double nature that also makes it hard to summarise. On one side, a free MIT-licensed engine that runs open-weight models on your own hardware without limits, without an API key and without a token bill. On the other, a metered cloud that takes over when a model outgrows your machine, using the same commands so nothing has to be rewritten. Few tools let you cross that line so cleanly, and the commitment never to log or train on prompts, stated three separate times across the legal pages and the pricing FAQ, is unusually firm for a hosted service.
The reservations sit on the corporate side rather than the product. Ollama Inc. publishes no postal address and no legal notice; a single mailbox handles support, legal, data protection and copyright; and the compliance toolkit a European buyer would expect is simply absent, with no data processing agreement, no subprocessor list, no Article 27 representative and no certification. The privacy policy also contradicts itself on the minimum age. None of this contradicts the privacy claims, but it does mean they have to be taken on trust rather than verified against documents.
The paid side is visibly under strain, too: new Max subscriptions are suspended for lack of capacity and the Team plan remains waitlisted, while usage allowances are described in words rather than in tokens. For an individual developer, the free local engine is close to unconditional. For an organisation buying the cloud, the questions to put to the vendor are already clear.
- Choosing a selection results in a full page refresh.
- Opens in a new window.