Is a Mac mini with Ollama enough for AI in your organisation?
For testing and a small team, yes: a Mac mini with Ollama runs medium-sized open AI models locally, without data leaving the building. For a whole organisation it is not enough. Memory limits model size and with it the quality of answers, Ollama handles one request after another by default, and sign-in, permissions, logs and backup are missing. As a guide, it suits one to five people.
What does a Mac mini with Ollama deliver?
Ollama is free software under the MIT licence that runs open AI models on a computer and makes them available to other programs through an interface. On a Mac mini it uses the memory shared by the processor and graphics. That is what makes the Mac mini popular: quiet, economical, quick to set up. For trying things out, for your own scripts or a small team it is a good solution, and the data stays on the device.
According to Apple, the Mac mini has at most 64 GB of unified memory (with M5 Pro, 307 GB/s memory bandwidth); the models with M6 have up to 32 GB (as of October 2026).
Which models fit on a Mac mini?
A rough calculation: for its weights, a model needs about as many gigabytes as it has billions of parameters, times bits per parameter divided by eight. On top of that comes memory for ongoing conversations and for the operating system.
| Model size | Weights at 4 bits | Fits on |
|---|---|---|
| 8 bn parameters | around 4 to 5 GB | any Mac mini |
| 30 bn parameters | around 15 to 18 GB | 32 GB and up, with room for conversations |
| 70 bn parameters | around 35 to 40 GB | only with 64 GB, just about |
| several hundred bn parameters | over 100 GB | no Mac mini |
A rough calculation, not a measurement. The best open models have several hundred billion parameters. So a Mac mini runs smaller models or heavily reduced versions, and the answers are weaker.
How fast does it answer?
When writing an answer, the computer reads the weights from memory for every token. Memory bandwidth therefore sets an upper limit: at 307 GB/s and a model with 18 GB of weights, that is at most around 17 tokens per second, or about ten words. For one person that feels fluent. Models that use only part of their parameters per token (mixture of experts) are faster. Reading in long documents takes noticeably longer on a Mac mini than on a server with GPUs.
How many people does it serve?
By default Ollama handles one request per model at a time; further requests wait in a queue (according to the documentation, up to 512). More parallelism can be configured, but then memory requirements grow with the number of concurrent requests times context length. In a test by Red Hat, Ollama’s performance on a server GPU barely increased with more people, while software built for many concurrent requests kept improving. As a guide, a Mac mini is enough for one to five people. How organisations grow beyond that is covered in the guide Scaling Ollama and Open WebUI.
What is missing for use across the organisation?
- Sign-in and permissions: Ollama’s interface has no sign-in of its own. By default it is only reachable on the device itself; if you open it up on the network, you have to secure access yourself. In 2025 Cisco found more than 1,100 Ollama servers that were openly reachable on the internet.
- User interface: Staff need a chat interface such as Open WebUI, which also has to be run and updated.
- Logs and evidence: Who asked what and when is not recorded. Data protection, the works council or staff council and information security need this.
- Backup and outages: A single device is a single point of failure.
- Updates: The operating system, Ollama, the interface and the models all need maintenance.
How does this work with basebox?
Technically, AI runs on such small computers for one to five people, with weaker answers. basebox does not offer this, because the AI there does not deliver what an organisation needs in everyday work. For small teams there is the cloud: from 5 people, without your own hardware, ready in 2 minutes. For larger organisations there are servers in Classes 1 to 3, which, depending on the class, answer 100, 150 or over 200 questions at the same moment. You will find the comparison with Ollama and local servers under basebox and Ollama.
Note: Memory and speed figures are rough calculations based on Apple’s technical specifications as of 11 October 2026, not measurements. A test with your own tasks is what gives reliable results.
Frequently asked questions
Does Ollama send data to the internet?
Not with locally run models; Ollama states that it does not see inputs in that case. However, Ollama also offers models in its own cloud. If you want to rule that out, switch off access to the cloud, according to the documentation with OLLAMA_NO_CLOUD=1 or a setting in the configuration.
May we put patient or citizen data on the Mac mini?
Technically, the data stays in the building. Legally, it needs the same measures as any other system: access protection, logs, backup, a record of processing activities and, depending on the use, a data protection impact assessment. A desktop computer without sign-in does not meet these requirements.
What about larger Apple computers?
Computers with more unified memory hold larger models. That changes little about the basic question: by default Ollama works through one request after another, and permissions, logs and operation do not come with it.
Sources
- Apple: Mac mini, technical specifications (accessed 11 October 2026, German) ↗
- Ollama: FAQ on concurrent requests, network and cloud ↗
- Ollama: licence (MIT, GitHub) ↗
- Cisco Talos: publicly reachable Ollama servers, analysis with Shodan (September 2025) ↗
- Red Hat: Ollama and vLLM under load compared (8 August 2025) ↗
This article was created with the help of AI, based on the sources listed. Responsible for the content: basebox GmbH. Not legal advice, no liability. Research as of 11 October 2026.