Home / Learn / Data & AI / Offline AI / Lesson 7.6
All lessons

Module 7: Getting and choosing models · Lesson 7.6 · about 15 min

Safe downloads and storage planning

How to download models without picking up malware or fakes, check that files arrived intact, and plan disk space for a growing model collection.

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

Module 7 · about 15 minutes · read first: 7.1 and 7.5. You’ll be able to build a safe, tidy model collection that fits your disk.

Get models from known publishers in data-only formats, check the fingerprint matches, and budget your disk like a cupboard you have to keep tidy.

1. Why this matters

Model files are large files that you download from the internet. Most are safe, but fakes and booby-trapped files exist. A few habits keep you safe and save disk space.

2. Safe habits

  1. Prefer known publishers: the model maker’s own account, or a converter with many downloads and a long history.
  2. Prefer data-only formats: GGUF and safetensors. Avoid .bin, .pt and .pkl from unknown sources (Lesson 7.5).
  3. Check the name: fakes copy famous names with small changes. Compare with the maker’s official page.
  4. Do not run scripts that come with a model unless you understand them; engines do not need them.
  5. Check the checksum for big manual downloads: model sites list a SHA256 for each file. On Windows certutil -hashfile file.gguf SHA256; on Mac or Linux shasum -a 256 file.gguf. Ollama checks this for you automatically.

A checksum is a fingerprint of a file: a 64-character code worked out from every byte. Change a single byte and the code comes out completely different. SHA256 is the name of the recipe used to make it.

Watch it happen: checking a download’s fingerprint

The published fingerprint. The model page lists the file's SHA256, which starts 3f9a… and ends …c71e, 64 characters in all.

Work out yours. After downloading, you run shasum -a 256 model.gguf (Mac or Linux) or certutil -hashfile model.gguf SHA256 (Windows). The computer reads all 7.7 GB and prints 64 characters.

Compare. They match exactly: 3f9a…c71e. Your file is the very file that was published, byte for byte.

What a mismatch means. If even one character differs, the file was damaged on the way or changed by someone. Delete it and download again from the maker's page.

“A checksum proves the model is safe.” It proves only that you got the exact file that was published. If the publisher uploaded a bad file, its checksum matches perfectly. That is why the publisher check comes first and the checksum second.

3. Storage planning

  • Gemma 4 E2B 4.6 GB, E4B 6.6 GB, 12B 7.7 GB, 31B 19 GB, gpt-oss-120b 65 GB (Ollama sizes).
  • Keep a short list: one small fast model, one strong model that fits, maybe one for code or images.
  • Delete what you no longer use: ollama list then ollama rm <name>; LM Studio has a My Models page.
  • Move the model folder to a bigger drive: Ollama uses the OLLAMA_MODELS setting; LM Studio lets you change its models directory in settings.
Maths and principle

Principle: trust the publisher, prefer data-only files, verify the bytes, and budget the disk.

Equation: disk needed ≈ sum of the model files you keep + 20% room for new versions.

disk≈1.2×∑file sizes

Analogy: buying medicine: from a known pharmacy, sealed packet, check the expiry date, and keep the cupboard tidy.

Try it: you keep E4B, 12B and 31B. Disk budget? ((6.6 + 7.7 + 19) × 1.2 ≈ 40 GB.)

Σ file sizes
the sizes of all the models you want to keep, added up
1.2
20% extra room, for a newer version you download before deleting the old one
SHA256
a 64-character fingerprint of a file; any change to the file changes it
Physical meaning
A model collection grows quietly: three models can fill 40 GB. Budget the space before you download, and keep only the models you actually use.

4. Worked example

Made-up numbers. Farah finds a “Gemma-4-31B-uncensored-FAST” upload from a two-day-old account, shipped as .pt files with a setup script. She skips it, pulls gemma4:31b with Ollama instead, and moves her models folder to a 1 TB external SSD.

5. Practice

  1. Run ollama list (or open LM Studio’s My Models). How much space do your models use?
  2. Check the SHA256 of one downloaded GGUF against the model page.
  3. Delete one model you have not used for a month.
Answers
  1. Add up the sizes; compare with your free space.
  2. If the two strings match exactly, the file is intact.
  3. You can always download it again.

New words: checksum (SHA256), publisher, model folder.

  • Download from the maker or a well-known converter, in GGUF or safetensors; skip pickle files and setup scripts from strangers.
  • A SHA256 checksum proves your file is exactly the published one; it does not prove the publisher is honest.
  • Disk budget ≈ 1.2 × the sum of the files you keep.
  • Keep a short list of models, delete unused ones, and move the model folder to a bigger drive if needed.

Check yourself

1Which download is safest?
2What does a checksum (SHA256) let you check?
3Where should models go on a laptop with a small internal drive?

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.