Ollama: run models on your own machine
Install a local model in one command, know what your hardware can run, and decide when local beats an API.
Pricing
Free and open source; you pay only for hardware (check the official page)
Official site
Open the official website (opens in a new tab)
Prices and plans change. Check the official page before you pay.
Ollama downloads and runs language models locally with a single command. It handles the model format, memory and a local API, which turns "run an LLM on my laptop" from a project into a step.
What it is good at
- Privacy. Nothing leaves the machine. For confidential documents, that settles the question.
- No per-token cost. After the hardware, usage is free, which matters for batch jobs over thousands of documents.
- Offline work. Flights, rural sites, air-gapped machines.
- A simple API. It exposes an HTTP endpoint compatible with common client libraries, so an app can switch between local and hosted models.
What it cannot do
A 7-billion-parameter model running on a laptop is not the same product as a frontier model. It is worse at reasoning chains, worse at long context, and more likely to invent. For summarising and extraction on well-defined text, the gap is small; for open-ended analysis, it is large.
Hardware is the limit. A 7B model needs about 8 GB of memory to run comfortably, a 13B model about 16 GB, and the largest open models need a serious GPU. Quantised versions trade a little quality for a lot of memory.
Getting started
Install the app, then run a model:
ollama run llama3.2 The first run downloads several gigabytes. After that, ollama list shows what is available, ollama pull fetches another, and ollama serve exposes the API on localhost.
Model choice is a real decision: general chat models, code models, embedding models and vision models are different downloads. Check the model card for size and licence before assuming it fits.
Where it fits in a real system
Local models are good for:
- Draft classification and extraction in a pipeline.
- A private assistant over internal documents.
- Development, so tests do not spend API credits.
Hosted models are better for anything user-facing where quality is visible.
Pricing
The software is free and open source. The cost is hardware, electricity and your time. Some models have licences restricting commercial use, so check each one.
Who it is for
Developers building offline or private features, organisations with data they cannot send anywhere, and anyone curious about how these models behave. It is not a drop-in replacement for a frontier assistant.
Alternatives
LM Studio offers a graphical interface over similar models. llama.cpp is the lower-level engine for people who want control. Hosted APIs remain the pragmatic choice when quality matters more than privacy.