Parameters: What "70B" Actually Means

Browse any AI announcement and you will see numbers like "8B", "70B", or "400B" attached to model names. People treat these like horsepower ratings, assuming a bigger number is always a better model. The truth is more interesting and more useful. That number counts the model's parameters, and understanding what a parameter is tells you why bigger is sometimes better, sometimes not, and why a small model on your laptop can still be genuinely useful.

This lesson keeps everything conceptual and no-code. By the end you will read "70B" and know exactly what it refers to and what tradeoffs come with it.

What You'll Learn

• What a parameter is, in plain language
• What "B" (billions) counts, and what it does not
• Why a bigger model is not automatically a better one
• Where small models win, and why they matter to you

A parameter is a learned setting

A parameter is a single adjustable number inside the model. You can picture the model as an enormous panel with billions of tiny dials. Each dial nudges how one signal flows into the next as text passes through. Training is the process of turning all those dials, little by little, until the model gets good at predicting the next token.

So "70B" means the model has roughly 70 billion of these dials, each holding a number that was learned from data. The parameters are the model's knowledge. There is no separate database of facts inside; everything the model "knows" is baked into the specific settings of those billions of dials.

A useful way to hold this:

• Parameters = the learned knobs, fixed once training ends. This is the "B" number.
• Tokens = the text flowing through those knobs at the moment you use it (from the earlier lessons).

They are completely different things, which is why a model can have a fixed parameter count but handle inputs of many different lengths.

Why more parameters can help

Broadly, more parameters give a model more capacity to store patterns and nuance. With more dials, a model can represent more facts, more languages, more subtle relationships, and longer chains of reasoning. This is why, all else equal, larger models have historically been more capable on hard, open-ended tasks.

But "all else equal" is doing a lot of work in that sentence, and in practice all else is rarely equal.

Why bigger is not always better

Parameter count is only one ingredient. Two others matter just as much:

• Training data quality and quantity. A model trained on more, cleaner, better-curated data can beat a larger model trained on less or messier data. Data is often the real bottleneck, not size.
• Training technique. Better methods, better fine-tuning, and better alignment (covered in a later lesson) can make a smaller model feel smarter than a clumsy larger one.

There is also a hard practical cost to size. A bigger model needs more memory and more computation to run, which means:

• It is slower to respond and produces tokens at a lower rate.
• It is more expensive to serve, because it uses more hardware per request.
• It may not fit on ordinary hardware at all, so you cannot run it privately on a laptop or phone.

Benchmarks and your own testing matter more than the headline number. A well-trained mid-size model can outperform a poorly trained larger one on the tasks you actually care about. The only reliable test is trying both on your real work.

Where small models win

Small models (the ones with single-digit or low-tens-of-billions parameter counts) are quietly one of the most important trends in AI, and they matter to everyday users for concrete reasons:

• They can run locally. A small model can fit on a good laptop or even a phone, which means it works offline, keeps your data on your device, and costs nothing per request. For privacy-sensitive or high-volume work, that is a big deal.
• They are fast and cheap. For narrow, well-defined jobs (classifying messages, extracting fields, simple rewriting, routing requests), a small model often does the job instantly at a fraction of the cost of a frontier model.
• They are easy to specialize. Because they are cheaper to fine-tune, teams can adapt a small model to one domain and get excellent results without frontier-scale budgets.

The practical mindset is right-sizing: match the model to the task. Reaching for the biggest available model for every job is like renting a moving truck to buy groceries. A lot of real-world AI work is better served by a small, fast, cheap model, with a large model reserved for the genuinely hard parts.

This "big versus small" tradeoff also sets up the next lesson. Model makers found a clever way to get much of the knowledge of a huge model while paying closer to a small model's running cost. That trick is called Mixture of Experts, and it is up next.

If you want the everyday-tools framing of choosing between models, the AI Essentials course and the FreeAcademy blog post on LLM parameters pair well with this lesson.

Key Takeaways

• A parameter is a single learned number inside the model, and "70B" means roughly 70 billion of them.
• Parameters hold the model's knowledge; there is no separate fact database inside.
• Parameters are fixed after training, while tokens are the text flowing through at use time, so the two are unrelated numbers.
• Bigger is not automatically better: data quality, training technique, speed, and cost all matter, and larger models are slower and pricier to run.
• Small models win when you need local, private, fast, cheap, or specialized performance, so right-size the model to the task.