The whole free stack, and what it costs you

For everyone without a corporate licence

A clinic with four staff. A school where the network blocks everything. A two-person accountancy practice. A charity. A co-operative. A firm in a country where the enterprise tier is either not sold or costs more per seat per month than a day's revenue.

This lesson is the whole free stack, named, with an honest account of what you lose.

The stack

Documents, spreadsheets, presentations, drawings, databases. LibreOffice — Writer, Calc, Impress, Draw and Base. Free, offline, on Windows, macOS and Linux, reads and writes Microsoft formats, does mail merge, pivot tables and PDF export.

The model itself. Ollama or LM Studio, both free, both a straightforward install. They download and run open-weight models on your own machine with no account and no network traffic. llama.cpp underneath them if you want it directly.

What runs on what, roughly: a 3-to-4-billion-parameter model on 8 GB of memory, a 7-to-8-billion model on 16 GB, larger with a graphics card or Apple silicon. On an ordinary CPU expect a few words per second — readable as it appears, slower than you are used to. Any decent GPU or an Apple M-series chip makes it comfortable.

Transcription. whisper.cpp runs Whisper offline; Vosk is lighter for modest hardware.

Text from documents. pdftotext from poppler-utils, ocrmypdf with Tesseract for scans, Tabula for tables in PDFs.

Search. Recoll or DocFetcher for indexed full-text search over your files; ripgrep for exact strings, instantly, over enormous folders.

Data. OpenRefine for cleaning, csvkit for inspecting, Python with pandas if somebody there writes a little code.

Mail. Thunderbird, with rules for the deterministic triage.

Images and design. GIMP, Krita, Inkscape, Scribus — covered properly in this site's creative courses.

Everything above is free, and most of it is also more private than the paid alternative, because it never connects to anything.

What you actually lose

The honest accounting, task by task. A local eight-billion-parameter model is:

Close to a frontier model on rewriting and tone, summarising a document you supply, classification with good examples, cleaning up dictation, structured extraction (especially with a grammar constraint), finding contradictions in a document, and drafting from your own notes.

Meaningfully worse on hard multi-step reasoning, long documents, code, unusual languages, and anything requiring broad world knowledge it was too small to hold.

Notice what that split means for this course. Most of the office work described in these seven blocks lands in the first list. The tasks in the second list are largely ones this course has been telling you to check anyway, or to avoid.

There is a second, less obvious advantage. You pin the version. A local model does not change under you on a Thursday because a provider shipped an update. For a process you have tested and documented, that stability is worth something real, and it is not available at any price from a cloud service.

The costs that are not zero

Setting it up takes an afternoon, and somebody has to be willing to. Somebody has to update it occasionally. The hardware is yours, and an eight-year-old laptop with 8 GB of memory will run a small model slowly and nothing larger. Electricity is not free, though at these scales it is negligible against any subscription.

And the interfaces are less polished. There is no audio overview, no beautifully integrated sidebar, no phone app that syncs.

The mixed policy most people should adopt

Very few organisations should be all-local or all-cloud. The workable rule is written down, in one page:

• Local, always, for anything containing personal data, client identifiers, patient information, staff matters, or commercially sensitive figures.
• Cloud, for public material, general drafting, research on published sources, and anything you would be content to see quoted.
• The test at the boundary: if you would hesitate to read it aloud to a stranger on a train, it goes local.

Write that down, put a name against it, and it becomes the policy your workplace did not have. The next block is about the duties that make this necessary rather than merely prudent.

Where to start if this is all new

Do not install nine things on a Monday. One a fortnight, each chosen because it removes a job you currently do by hand.

A reasonable order for most small organisations: LibreOffice first, because it replaces the largest recurring cost and everything else works alongside it. Then Recoll or ripgrep, because finding documents is the complaint you hear most often. Then ocrmypdf, if paper arrives. Then Ollama, last, because it is the one that needs somebody to keep an eye on it and the one whose value depends on the others being in place.

And keep one commercial account, if you can afford a single seat. Having a frontier model available for the hard questions — with nothing confidential in them — makes the local stack easier to defend, because nobody has to argue that the free tools are as good at everything. They are not, and saying so is what makes the rest of the argument credible.