QUICK ANSWER:
Ollama is free software that downloads and runs open-weight AI models on your own computer with a single command. Nothing leaves the machine, which is as strong as the privacy position gets for an Australian business. The one thing to size up front is video memory: at four-bit quantisation you need roughly 0.6 to 0.7 GB of VRAM per billion parameters, so a 7B model wants about 4 GB, a 13B wants about 8 GB, and a 70B needs 40 GB or more.
Every other tool in this series sends your data somewhere. Ollama is the exception, and that single fact makes it worth understanding even if you never install it, because it is the benchmark you should be measuring the others against.
What Ollama does
Ollama is the only tool in this category where the honest answer to what happens to your data is nothing. It never leaves the machine, because there is no server on the other end.
Ollama is a small application that handles the unglamorous parts of running an AI model on your own hardware: downloading it, quantising it to fit, serving it over a local API and swapping between models. You type one command, it pulls the model, and you have a working assistant offline.
The part that matters for business use is that it exposes an OpenAI-compatible endpoint on your own machine. Most tools that expect to talk to a commercial AI provider can be pointed at Ollama instead by changing one address. That is how you take a workflow that was sending client documents overseas and make it stop, without rebuilding the workflow.
The hardware question, answered properly
This is where most guides wave their hands. Here is the arithmetic.
Models are usually run at four-bit quantisation, which is the sweet spot between quality and memory. At that setting, budget roughly 0.6 to 0.7 GB of video memory per billion parameters, then add headroom for the context window. That gives you a usable table:
| Model size | VRAM needed (4-bit) | Realistic hardware | What it is good for |
|---|---|---|---|
| 3B to 8B | About 4 to 6 GB | A recent laptop GPU, or an Apple silicon Mac | Summarising, drafting, classification, extraction |
| 13B | About 8 to 10 GB | A mid-range desktop GPU | Better reasoning, longer documents |
| 30B class | About 20 to 24 GB | A high-end desktop card | Work you would otherwise send to a paid API |
| 70B and above | 40 GB and up | Server hardware or a rented cloud GPU | Frontier-adjacent quality, at frontier-adjacent cost |
One thing worth knowing before you buy anything: unified memory on Apple silicon behaves differently from discrete VRAM, and a Mac with 32 GB of shared memory will happily run models that would need a very expensive PC card. If you are choosing hardware specifically for this, that changes the answer.
What it costs
There is also a hosted option. Ollama Cloud runs the same models on their GPUs and keeps the same interface, so your local setup works unchanged against a cloud model. There is a free tier behind daily quotas, a single-user paid tier that opens the full catalogue and higher limits, and larger team and production tiers above it. All of them are flat monthly subscriptions rather than per-token bills, which makes them easy to budget against, and all of them are priced in US dollars, so add currency conversion and GST where it applies.
Worth matching to your reason for choosing Ollama: the Cloud tier gives you the same interface without the hardware, which is ideal for testing and for bigger models. Keep the work local when privacy is what you came for, and you get the best of both.
Why this matters more in Australia than the coverage suggests
Australian Privacy Principle 8 covers cross-border disclosure. The moment personal information leaves Australia to an overseas recipient, you carry the obligation, and if that recipient handles it in a way that would breach the Australian Privacy Principles, you are treated as having breached them yourself. You cannot delegate that by pointing at a vendor’s privacy policy.
A local model is the only architecture that sidesteps the question entirely, because there is no disclosure. That is a genuinely strong position for anyone handling health information, financial detail, tenancy applications, employment records or anything else you would prefer to keep entirely in-house.
The second reason is the automated decision-making transparency requirements under the Privacy Act, which cover decisions that significantly affect people. Running the model locally does not exempt you from disclosing those, but it does mean you can point to exactly where the processing happened and what data it saw, which is a far easier conversation than reconstructing it from a vendor’s logs.
How to actually get value from it
Local models pay off fastest when you match them to the right work. Think of one as a different tool rather than a cheaper substitute for a frontier model, and point it at the jobs that are high-volume and low-judgement. Do that and it earns its keep in the first week.
Good fits: turning meeting recordings into structured notes, extracting line items from supplier documents, categorising inbound enquiries before a human sees them, redacting personal information before it goes to a hosted model, first-pass drafting that a person will edit anyway.
Leave to a hosted model: anything where a subtle mistake is costly, complex multi-step reasoning, and client-facing work where the extra polish shows. Running both is the setup most businesses land on, and each one does what it is best at.
The pattern we recommend most often is a two-tier setup. A local model does the bulk work and the redaction, and a hosted model handles the small fraction of tasks that genuinely need more capability. It cuts the API bill sharply and shrinks the amount of sensitive data that ever leaves the building. Pairing Ollama with an agent like OpenClaw is the usual way that gets wired together.
Which model to run
Ollama is the engine, not the model. The interesting choices right now are the open-weight families covered elsewhere in this series: Qwen for general work and vision, Mistral for fast European-built models, and the smaller distilled models when hardware is tight. Check the licence on whichever you pick, because open weights and open licences are not the same thing, and a couple of the most popular families have commercial conditions attached.
Pros and Cons
Pros
- Free, unlimited and completely private on your own hardware.
- No cross-border disclosure, which takes an entire compliance question off the table.
- Works offline, including on sites with poor connectivity.
- OpenAI-compatible endpoint, so existing tools can point at it with one change.
- Predictable cost, because there is no meter running.
Worth knowing before you start
- Best on high-volume routine work; frontier models still lead on hard reasoning.
- The largest models want server-class hardware or a rented cloud GPU.
- Uses the machine while it generates, so pick where it runs.
- Someone has to keep models and the runtime updated.
- Model licences vary, and some restrict commercial use.
The verdict
Ollama is the most underrated tool on this list for Australian businesses specifically, because the compliance advantage is worth more than the capability gap costs. If you handle personal information and you have been nervous about pasting it into a hosted chatbot, this is the answer to that nervousness, and it costs nothing to try on the laptop you are reading this on.
Start with a 7B or 8B model on existing hardware, give it one boring high-volume job, and measure whether the output is good enough. If it is, you have removed a recurring cost and a recurring risk in the same move.
Work out what should stay in-house
The useful question is not whether to run AI locally, it is which parts of your workflow should never leave the building and which are fine in the cloud. That split is what our AI consulting engagements map out first, before any tool gets chosen. If you want your team to be able to make that call themselves, our AI training covers the same ground in a workshop.
It is free, and it runs on the machine you are reading this on. You can find it at Ollama.
Frequently asked questions
Is Ollama free?
Running models locally with Ollama is free and unlimited. There is no per-token charge and no seat licence. Ollama Cloud is a separate paid option, with a free quota-limited tier, a single-user paid tier and larger team and production tiers above it, all billed as flat monthly subscriptions.
How much VRAM do I need to run Ollama?
At four-bit quantisation, budget about 0.6 to 0.7 GB of VRAM per billion parameters plus context headroom. That means roughly 4 GB for a 7B model, 8 GB for a 13B, and 40 GB or more for a 70B. Apple silicon uses unified memory and can run larger models than the equivalent PC VRAM figure suggests.
Does Ollama send my data anywhere?
No. When you run a model locally, the text never leaves your machine. That is the main reason to choose it over a hosted API, and it means there is no cross-border disclosure to manage under Australian Privacy Principle 8.
Is a local model good enough for real work?
For high-volume, low-judgement work such as summarising, extraction, classification and first drafts, yes. For difficult reasoning, long multi-step tasks or anything client-facing where quality is obvious, a frontier hosted model is still noticeably better.
Can Ollama replace my ChatGPT subscription?
Partly. It can absolutely replace the routine drafting and summarising you currently paste into a chatbot, especially anything containing personal or commercial information. It will not match a frontier model on the hardest tasks, so most businesses end up running both.
