Ollama
Authorship and approval
[Written with AI] [Approved by a human]
Updated
03.11.2026
Ollama is a free, open-source program that downloads and runs language models on your computer. The model runs locally: text never leaves the cloud, and you don't pay for prompts.
6.2 GB
9B model RAM on our setup
3.0 tok/s
9B model on a 2011 CPU
19.5 h
2k texts tagged locally
How does it work?
- ollama pull downloads the model, usually pre-reduced (quantization) to fit on a standard computer.
- ollama run loads the model into memory and lets you chat with it in the terminal.
- Ollama also exposes a local API, so other programs can use the model, such as n8n automations or your own script.
- After the last prompt, the model stays in memory for a few more minutes (5 by default) so the next prompt doesn't wait for a reload. ollama ps shows what is loaded, and ollama stop frees memory immediately.
How much RAM is needed?
- Rule of thumb: a model needs roughly as much RAM as its file size, plus headroom for the conversation (context window).
- 8 GB RAM: comfortably runs ~4B models (B stands for billions of parameters). 16 GB RAM: models up to ~11B.
- A graphics card is not required. Without one, the model runs on the CPU and responds more slowly; with one, it is several times faster.
- When RAM runs out, the system swaps memory to the disk (swap) and everything slows down drastically, not just the model.
How to choose a model?
- Task: a small model is enough for sorting and tagging text. For conversation, writing, and difficult questions, a larger one is needed.
- Language: small models write in Polish worse than in English. Polish models (like Bielik) are trained specifically with the Polish language in mind.
- Memory: first check how much RAM you have, then look for the largest model that fits with headroom.
- License: Ollama is free, but every model has its own license. You must read it for commercial use.
How our lab does it
The lab server is an Intel i5-2400 from 2011, with 15 GB RAM and no graphics card. In November 2026, we benchmarked the same prompts across several models:
- Qwen 3.5 4B: 3.1 GB RAM, 5.0 tokens per second.
- Ornith 1.5 9B: 6.2 GB RAM, 3.0 tokens per second.
- Bielik 11B: 7.7 GB RAM, 2.2 tokens per second.
- Ministral 3 14B: 9.1 GB RAM, 1.5 tokens per second. Too slow for chat, fine for overnight batch jobs.
Everyday, Ollama uses our setup to tag research papers for our custom topic classifier: Qwen 3.5 4B tagged 2k texts in 19.5 h, in the background and at zero cost. Lesson: pausing the program sending prompts didn't free up memory because Ollama was still finishing the previous response. The server hit swap. Only ollama stop brought RAM usage down from 12 to 3 GB.
Related terms
Model quantization · Large Language Model (LLM) · n8n · Tokens and Context Window
Questions and answers
Is Ollama free?
Yes, it is open-source software under the MIT license. Models also have their own licenses which must be checked individually, especially for commercial use.
Do I need a graphics card for Ollama?
No. The model also runs on the CPU alone, just slower. All our benchmarks come from a computer without a GPU.
Will a local model replace ChatGPT?
For simple, repetitive tasks, often yes. For difficult prompts and high-quality Polish writing, large cloud models are still significantly better today.
Sources
- Ollama, repository and documentation (MIT license)
- Lab: model benchmark on i5-2400 and B138 topic classifier measurements (November 2026)
Next