Home / Learn / Data & AI / Offline AI / Lesson 7.1
All lessons

Module 7: Getting and choosing models · Lesson 7.1 · about 20 min

Where models live

Hugging Face, the Ollama library and the catalogues inside desktop apps. Where to find models, how to read a model page, and who to trust.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 7 · about 20 minutes · read first: 1.2. You’ll be able to find a model, read its page and judge whether to trust the download.

Models live on a few big public shelves; pick a copy from a maker you can name, with a clear page and many users, and you avoid almost every bad download.

1. Why this matters

There are hundreds of thousands of models online. The popular ones have many copies, made by many different people. Some copies are great. Some are broken. A few are unsafe.

Knowing where to look, and what to check, saves you hours and keeps bad files off your computer.

2. The three main shelves

Think of a model as a book. These are the libraries that lend it out:

Where What you find Best for
Hugging Face (huggingface.co) Original models from their makers, plus converted versions (GGUF, MLX) by others Choice, details, any engine
Ollama library (ollama.com/library) Ready-to-run models with simple tags such as gemma4:12b Ollama users, quick start
App catalogues (LM Studio and others) A search inside the app, usually drawing on Hugging Face Clicking instead of typing

Some makers also publish on their own sites or on Kaggle.

A publisher is whoever uploaded a copy. A conversion is the same model saved in another file format (Lesson 7.5) or shrunk to a smaller size (Module 8). The model maker is one publisher; people who convert models are others.

3. Reading a model page

Every good model page is called a model card. It is like the label on a medicine box: who made it, what is inside, and what it is for. A model card tells you:

  • who made it and when;
  • size: parameters, and for MoE models, active parameters;
  • context length;
  • licence (Lesson 7.2): what you may and may not do with it;
  • intended use and limits, languages, and known weaknesses;
  • files: the formats and quantisations available, with sizes.

If a page has none of this, skip it.

4. Who to trust

Prefer, in this order:

  1. the original maker (for example Google for Gemma, Alibaba’s Qwen team, OpenAI for gpt-oss);
  2. well-known converters whose files are widely downloaded;
  3. everyone else, only if you have a good reason.

Lesson 7.6 covers checking a download once you have it.

Watch it happen: picking one download from 300

Search. You type "gemma 4" on Hugging Face and get about 300 results. Most are copies, fine-tunes and experiments by other people.

Start at the maker. You open Google's own Gemma 4 page. It names the maker, the sizes, the context length and the licence (Apache 2.0). Now you know what the real model is.

Pick a format. Your app needs a GGUF file. The maker's page links to conversions; you look for a publisher you have heard of.

Check the signals. Three candidate copies:

CopyLicenceFormatDownloads
AnamedGGUF1,200,000
BnoneGGUF40
Cnamed.pt9,000

Only A passes every check.

Choose the size. Inside copy A you pick the Q4_K_M file that fits your memory (Lesson 8.5). Then you download it once and keep it.

Maths and principle

Principle: trust grows with independent checks: a known maker, many users, and matching file fingerprints.

Equation: risk ≈ chance of a bad file × how much the file can do on your machine.

Analogy: buying medicine. A pharmacy brand with a sealed box beats an unlabelled bottle from a stranger, even if the label says the same thing.

Try it: two copies of a model, one downloaded 2 million times from the maker, one 40 times from an unknown account. Which first? (The maker’s.)

Chance of a bad file
how likely this copy is broken, fake or booby-trapped. Low for the maker's own page, higher for unknown accounts.
What the file can do
data-only files (GGUF, safetensors) can only hold numbers; older pickle files can run code when opened.
Risk
the two multiplied: a safe format from a known maker is close to zero.
Physical meaning
Every check you add (maker, licence, downloads, format) cuts the chance of trouble. Thousands of other people using the same file have, in effect, tested it for you.

“Newer upload, better model.” A copy uploaded an hour ago is not an improved model; often it is a re-upload with a catchy name and no details. The model itself only changes when the maker releases a new version, and the maker’s page will say so.

5. Worked example

Made-up numbers. Kwame searches “gemma 4” on Hugging Face and finds 300 results. He filters to the maker’s own page, reads the card (Apache 2.0 licence, 256K context for the 31B), then picks a GGUF conversion by a well-known publisher with over a million downloads, at Q4_K_M.

How long will it take? Model files are big, so the download time is worth knowing. A 7.7 GB file on a 100 Mbit/s line: 7.7 × 8,000 = 61,600 megabits, ÷ 100 = 616 seconds, about 10 minutes.

6. Practice

  1. Open the Ollama library page for gemma4 and list its tags.
  2. Open a Gemma 4 model card on Hugging Face. Find the licence and context length.
  3. Name two warning signs on a model page.
Answers
  1. You should see sizes such as e2b, e4b, 12b, 26b-a4b and 31b.
  2. Licence and context are near the top of most cards.
  3. No maker or licence named; very few downloads; a name copying a famous model; files in formats that can run code (Lesson 7.6).

New words: model card (a model’s description page), publisher (who uploaded it), conversion (the same model in another file format).

  • Three main shelves: Hugging Face (most choice), the Ollama library (ready to run) and the catalogue inside your desktop app.
  • A model card says who made the model, its size, context, licence, limits and files. No card, no download.
  • Trust in order: the original maker, then well-known converters with many downloads, then others only with good reason.
  • Download time in minutes ≈ GB × 8,000 ÷ your speed in Mbit/s ÷ 60.

Check yourself

1What is Hugging Face?
2Two downloads of "Gemma 4 12B" come from different publishers. What should you prefer?
3Where does Ollama get the model when you run ollama pull gemma4:12b?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.