Special Topic · Open Weights, Distillation & Local Runs

What Are Weights? Everything a Model Knows How to Do

The file that months of training finally condenses into: what it looks like, how big it is, and why holding the weights means holding control

THE QUESTION THIS PAGE ANSWERS

ANSWER FIRST

What Are Weights? Everything a Model Knows How to Do?

The file that months of training finally condenses into: what it looks like, how big it is, and why holding the weights means holding control

DECISION RULE

Make the claim earn its place. Use this page as a decision aid, not a definition to memorize. Connect the idea to one real task, one observable result, and one failure that would change your mind.

TRY NEXT

Write one question you could answer with evidence after trying this idea.

WATCH FOR

A conclusion that sounds complete but leaves the key assumption untested.

The short answer
Weights are the pile of numbers a model is left with after training. The tens of millions of dollars and the months of compute all end up condensed into this one file. Get hold of it and you can run exactly the same model on your own machine — no network connection, no payment to anyone.

Almost every example in this chapter comes from Qwen, so let us say why up front.

  • It covers the size range most completely. In the official Ollama library the Qwen3.5 generation runs from 0.8B all the way up to 122B, with 2B, 4B, 9B, 27B and 35B in between. Every hardware tier discussed later maps onto a file you can genuinely download, without switching brands to fill the gaps.
  • Other people build on top of it. Four of the six distilled versions DeepSeek released for R1 use a Qwen base — that is a third party's choice, not a vendor's own marketing.
  • It also serves as its own cautionary tale. During the Qwen2.5 era the 3B and 72B models shipped under a non-standard license, which makes it a useful illustration of how unreliable the word "open source" can be. That discussion is in lesson 2.

In the Chinese-speaking community some people call it "the god of open source". That is a nickname, not a rating, and it does not mean Qwen leads on every dimension. This chapter uses it for examples because the material is complete and easy to work through, not because it is the best. What you should walk away with is a set of criteria, not a brand name — the same criteria should work just as well applied to any other vendor.

What it looks like

There are no rules inside a model, no knowledge base, no if-else. There is only an enormous table of numbers. Every number in the cells below is one weight, and the darker the color, the larger its absolute value.

Positive Negative Darker means a larger absolute value

That is 36 numbers. A Qwen3-8B model has 8 billion of them. The entire point of training is to nudge those billions of numbers from random values toward the right ones. Once that is done, the parameters are frozen and the model is fixed.

This also explains something from Chapter One: a model learns nothing while it is chatting with you. Weights are fixed during training; during a conversation they are read-only. Whatever you tell it is gone by next time.
Parameter count determines file size

How many bytes each weight takes depends on the storage precision. Models are usually released in FP16, which means 2 bytes per number. That gives you an easy estimate for file size:

Weight file size (GB) ≈ parameter count (B) × 2
8 billion parameters × 2 bytes = 16 billion bytes ≈ 16 GB

You can memorize the formula, but what it means only lands once you drag it yourself. The slider below runs from 0.5B up to 2.4 trillion, and you can switch storage precision as you go to see the same model's file swell and shrink.

Drag the parameter count to 8B, then switch once between FP16 and INT4.
Parameters 8B
Precision
16 GB

Here is the comparison across a set of real models. Qwen serves as the yardstick because its size lineup is currently the most complete — public releases run from 0.6B all the way to 2.4T — so comparing within one family keeps architectural differences out of the picture.

Qwen3-0.6Babout 1.2 GB
Qwen3-8Babout 16 GB
Qwen3-32Babout 64 GB
Qwen3-235B-A22Babout 470 GB
Qwen3.8-Max (2.4T parameters)several TB
Parameter counts come from each model's official release notes; sizes are theoretical FP16 estimates, and actual files differ slightly depending on how they are sharded and which config files ship alongside them. Qwen3.8-Max has 2.4T total parameters with 95B active, released 2026-08-03.
One thing people mix up

Storage size and runtime footprint are two different things. A 16 GB file does not mean 16 GB of VRAM will run it. At runtime the model also needs room to hold the intermediate state of the conversation — the KV Cache from Chapter Two. In practice, budget another 30% to 50% on top of the parameter size.

There is an interactive tool at the end of this chapter where you can pick your own GPU or Mac model and see what will run. For now, keep the conclusion: estimating VRAM from file size will leave you short.

Why weights matter so much

Put all of that together and the significance of weights is clear. They are not just a file. They are control.

Offline
No dependence on any provider
It runs with the network cable unplugged. Price hikes, rate limits, and deprecations cannot touch you
Editable
You can keep training on top of it
Fine-tune your own version on your own data, which is something an API call cannot do
Private
Data never leaves your machine
For medical records, legal files, and internal code that cannot go out the door, this is the only path

Turn that around: with an API-only model, what you rent is the right to use it. Price, availability, and the retirement date are all decided by the other side. During technology selection that difference turns directly into risk, and it is what the next section takes apart: "open source" means very different things depending on who is saying it.

Put “The short answer” back into its constraints

“Almost every example in this chapter comes from Qwen, so let us say why up front” shows that a model, license, access route, or leaderboard is information—not an answer outside context. The real choice depends on task, data boundary, latency, quality floor, and operating cost.

Write elimination criteria before chasing the top score

The comparison in “In the Chinese-speaking community some people call it "the god of open source" .” should use the same real inputs while observing correctness, failure behavior, response time, and cost. A model leading a public leaderboard may still fail your license, privacy, or peak-latency constraints.

  • Other people build on top of it. Four of the six distilled versions DeepSeek released for R1 use a Qwen base — that is a third party's choice, not a vendor's own marketing

Without a test set, there is no reliable winner

Start with “Turn that around: with an API-only model, what you rent is the right to use it.”: choose inputs that could genuinely change the decision and write down one counterexample that would reverse your choice. That is more useful than memorizing a single ranking.

From “The short answer” to “What it looks like”

“The short answer” grounds the problem in “Almost every example in this chapter comes from Qwen, so let us say why up front”. “What it looks like” then moves it toward “There are no rules inside a model, no knowledge base, no if-else. There is only an enormous table of numbers. Every number in the cells below is one weight, and the darker the color, the larger its absolute val…”. Together, they show that the lesson is not just a conclusion to remember, but a claim with conditions.

Carry the judgment into the next situation

For model selection, write non-negotiable constraints from the real task first. Compare quality, failure behavior, latency, licensing, and cost on the same inputs; use a leaderboard only as a starting point.

  • “The short answer”: Almost every example in this chapter comes from Qwen, so let us say why up front
  • “What it looks like”: There are no rules inside a model, no knowledge base, no if-else. There is only an enormous table of numbers. Every number in the cells below is one weight, and the darker the color, the larger its absolute val…
  • “The closing point”: Other people build on top of it. Four of the six distilled versions DeepSeek released for R1 use a Qwen base — that is a third party's choice, not a vendor's own marketing

The final “The closing point” brings the discussion to “Other people build on top of it. Four of the six distilled versions DeepSeek released for R1 use a Qwen base — that is a third party's choice, not a vendor's own marketing”. The useful thing to carry forward is knowing which judgments must be revisited when input, scale, or risk changes.

Mark as learned Your reading progress updates automatically
← PreviousNext →

Keep reading

The next useful article in the thread.

ARTICLE DISCUSSION

Leave one useful thought here.

Keep the idea that clicked, the question that stayed open, or a small note for the next learner.

Discussing What Are Weights? Everything a Model Knows How to Do Open Weights, Distillation & Local Runs
3discussionsArticle discussion · synced with the Circle
View in the learning circle
AM
Asha MorganContent editor
INSIGHTField note

I turned one judgment from this article into a small experiment I could run today. Knowing what to observe next is more useful than simply remembering the conclusion.

ARTICLE DISCUSSION7 helpful
LH
Lin HarperIndie developer
INSIGHTInsight

After reading this, I first looked for the conditions behind the idea instead of copying the method into a project. That order made the later trade-offs much clearer.

ARTICLE DISCUSSION5 helpful
KM
Kiki MooreProduct operations
QUESTIONQuestion

When this judgment reaches real work, which constraint should be added first? I am curious which step matters most between reading and the first practical attempt.

ARTICLE DISCUSSION4 helpful