Ollama is an open-source program that downloads open AI models and runs them on your own machine, without sending the text to a cloud provider. It makes sense when data privacy or a predictable cost matters more than top quality. It does not make sense if you want the best possible answer with the least work: for that, a cloud API is still the simplest road.
It is a third-party project, installed on a VPS with root access. Interweb provides the server; installing and maintaining Ollama and the models is up to you. The install steps and the available models are on the official Ollama website, which changes often.
Cloud or on your own server
|
Cloud API |
Ollama on your VPS |
| Where the text goes |
To the model provider |
Stays on your machine |
| Cost |
Per use, grows with consumption |
The server, plus the work of keeping it. The per-request charge disappears. |
| Quality |
Access to the biggest and newest models |
Limited to the models your machine can handle |
| Speed |
Fast and steady |
Depends on CPU and memory. Without a graphics card, it can be slow. |
| Upkeep |
None on your side |
Updating the program, the models and the system |
When it is worth it
| 1 |
Data that must not leave. Internal documents, contracts, customer lists. If company policy or the law requires it, running on your own server avoids sending them to third parties. See personal data protection.
|
|
| 2 |
High volume, simple tasks. Sorting messages or summarising short texts, every day, at a fixed cost.
|
|
| 3 |
Experimenting without paying per request. For learning, a small model on a test VPS gives you the essentials.
|
|
| 4 |
An agent that should depend on nobody. OpenClaw can use local models, but that is where the strong machine is demanded. See what it is and what you need.
|
|
What decides whether your server copes
What rules is the size of the model. A bigger model needs more memory, and if it does not fit in the server’s memory it will not run, or will run so slowly it is useless. Each model states its memory requirements on its page; read them before you download it. To see what your VPS has, use free -h and nproc. Plan specifications are on the VPS, Cloud and Dedicated pages; check there whether any includes a graphics card, because without one the models run on the CPU.
|
Do not open Ollama to the internet. By default it usually listens only on the machine itself; confirm that in the documentation. If you expose it unprotected, anyone can use your server to run models. See ports and firewall on a VPS.
|
|
A local model makes things up too. Running on your own server protects the data, it does not guarantee right answers. See AI hallucinations.
|
|
Start with a small model and measure: how long it takes, how much memory it uses, whether the quality is enough for the task. Move up in size only if the result justifies it.
|
|
Want a server with more memory to try local models? Compare the plans.
See VPS plans
|
RECOMMENDED PRODUCT VPS server with root access Resources of your own, the OS you choose, reinstall whenever you like. from $7.61/mo (3-year plan, with coupon) See plans |
Português
English
