TPEE Toptronic® Prompt Engineering Editor

102. Model Weights 1 - Getting Started

Know what a weight actually is, why it is frozen, and how parameter count becomes gigabytes on disk.

By Jacques Botte, founder of Toptronic®. Last updated 17 October 2026.

The lesson

A weight is a single number a model learned during training. Nobody wrote it and nobody can read it as a sentence. Collectively the weights are everything the model knows: the grammar of language, the shape of code, the associations between concepts. The parameter count on a model's label, 7 billion or 70 billion, is simply how many of these numbers there are.

Weights are set once and then frozen. Using a model does not change them, no matter what you type, correct or explain. This is why a model forgets your correction between sessions and why it cannot learn your codebase by talking to you. The only thing that changes at use time is the context you supply, which the frozen weights are applied to. If you remember one sentence from this lesson, make it that one.

Size follows directly from arithmetic. Each weight is stored in a fixed number of bytes, so file size is approximately parameter count times bytes per weight, plus a small overhead for the tokenizer and metadata. That is why a 7B model is roughly 14 GB in 16-bit form and 4 GB in 4-bit form, and why 'can my machine run this' is a question about bytes, not about intelligence.

The words around weights are worth separating. Open weights means the numbers are published and you may download them; it says nothing about the training data or the code, and it usually comes with a licence that restricts commercial use or requires attribution. Open source, in the strict sense, would include data and training code, which very few models do. 'Free to download' and 'free to run' are different claims again: hosting weights on a server costs real money.

Finally, weights carry provenance risk. A downloaded checkpoint is a binary blob that a loader will deserialize, and old formats could execute code on load. Prefer the safetensors format, prefer a source you can name, and verify the file's hash when one is published. This app never downloads weights; the discipline is for the separate tools you use, and it is the same discipline you already apply to any binary you install.

Check yourself

Question 1: What is a model weight?
  1. A rule written by the vendor
  2. A number learned during training — correct
  3. A line of source code
  4. A token in the vocabulary

Answer: A number learned during training

Weights are learned numbers. Their meaning is distributed across the whole set and cannot be read individually.

Question 2: Why does a model not learn from your corrections during a session?
  1. Because the session is too short
  2. Because weights are frozen at inference; only the context changes — correct
  3. Because corrections are stored elsewhere
  4. Because the tokenizer discards them

Answer: Because weights are frozen at inference; only the context changes

Inference never writes back to the weights. Continuity must come from context or files outside the model.

Question 3: Why is safetensors preferred over older checkpoint formats?
  1. It compresses better
  2. It stores tensors without a deserializer that can execute code — correct
  3. It is the only format GPUs read
  4. It contains the licence

Answer: It stores tensors without a deserializer that can execute code

Older formats could run code at load time; safetensors is a pure data format, which removes that class of risk.

← Previous lesson · All 106 lessons · Next lesson →

The full course — 106 lessons and 318 quiz questions — ships inside the app. Get TPEE to study it offline.