AI Model Parameters Explained: 7B, 70B and More

AI model parameters are the numbers a neural network learns during training, and the “7B” or “70B” in a model’s name counts them in billions. Each parameter is a small value, usually called a weight, that nudges how the model turns one piece of text into the next. More parameters give a model more room to store patterns, but size alone does not decide quality, speed or cost.

This guide explains what AI model parameters are, how to turn a parameter count into a memory estimate, why a bigger model is not always a better one, how mixture-of-experts changes the maths and what the number means when you choose an assistant. It uses plain words and a few simple calculations, and no vendor-specific figures.

What AI model parameters actually are

A language model is a very large mathematical function. Text goes in as tokens, passes through many layers of arithmetic and comes out as a probability for every possible next token. The behaviour of that function is set by its parameters: long lists of numbers that decide how strongly each connection in the network responds. Training is the process of adjusting those numbers, billions of times, until the model predicts text well.

It helps to be clear about what parameters are not. They are not a database of facts that the model looks up, and they are not lines of code. Knowledge is spread across the numbers in a way nobody can read directly. This is also why a model can state something confidently and wrongly, which our guide to AI hallucinations covers. For the architecture behind the layers, see transformer models explained.

Parameters are also different from other numbers people mention:

Hyperparameters are settings chosen by the builders before training, such as the number of layers or the learning rate. They are not learned.
Tokens are the pieces of text the model reads and writes. They are not the same thing as words.
The context window is how much text the model can consider at once. It is a separate limit, covered in our guide to context windows.
Training tokens are how much text the model learned from, not how big the model is.

Reading “7B”, “70B” and “400B”

The letter B stands for billion. A 7B model has about seven billion parameters, a 70B model about seventy billion. The numbers are rounded labels, and two models with the same label can differ in design. Some providers do not publish counts at all, particularly for closed models, so you often cannot compare them on this number.

Think of the count as a rough size class:

Small (roughly under 10 billion): fast and cheap to run, good at simple tasks, can run on a laptop or phone with the right setup.
Medium (tens of billions): a good balance for many professional tasks.
Large (hundreds of billions and above): strongest at complex reasoning and long, difficult tasks, and the most expensive to serve.

These bands are loose and move as methods improve. A well-trained small model today can beat a poorly trained larger one from a few years ago, as our article on small language models explains.

From parameter count to memory

You can estimate how much memory a model’s weights need with one multiplication: parameters times bytes per parameter. Models are commonly stored in 16-bit numbers (2 bytes), and compressed versions use 8 bits (1 byte) or 4 bits (half a byte). The method of shrinking numbers is called quantization, described in our guide to AI quantization.

Parameters	16-bit weights	8-bit weights	4-bit weights
7 billion	about 14 GB	about 7 GB	about 3.5 GB
13 billion	about 26 GB	about 13 GB	about 6.5 GB
70 billion	about 140 GB	about 70 GB	about 35 GB
400 billion	about 800 GB	about 400 GB	about 200 GB

These are weights only. Running the model also needs memory for the conversation, which grows with the context length, and for overhead. In practice you need noticeably more than the table shows. The table does explain why models with hundreds of billions of parameters run on racks of specialised hardware, while a 7B model can run on a good home computer.

Do more parameters mean a better model?

Often, but not always. Research on scaling has repeatedly found that, with enough good data and training, larger models get better at many tasks. Yet size is one input among several:

Training data quality and amount. A well-known study from 2022, often called the Chinchilla paper, found that many large models had been trained on too little text for their size, and that a smaller model trained on more data could match a larger one. You can read it on arXiv.
Training after pre-training. Instruction tuning and feedback, such as learning from human preference rankings, change how helpful and safe a model feels without changing its size.
Architecture. Newer designs get more from each parameter.
Specialisation. A model tuned for code or translation can beat a larger general one on that task.
Reasoning methods. Letting a model think longer at answer time can raise accuracy without more parameters, as in reasoning models.

The honest summary is that parameter count predicts capacity and cost better than it predicts quality on your task. Benchmarks and your own tests are better guides. Our article on how to read AI benchmarks shows how to use them without being misled.

Total versus active parameters in mixture-of-experts models

Some large models use a design called mixture of experts. The network contains many specialist sub-networks, and for each token only a few of them run. The model then has a large total parameter count but a much smaller number of active parameters per token. Total parameters tell you how much knowledge capacity and memory the model needs. Active parameters tell you roughly how much computation each token costs.

This is why a very large mixture-of-experts model can answer quickly. Its memory footprint is large, but each step uses only part of it. When a provider quotes two numbers, such as a larger total and a smaller active figure, that is what they mean. Our explainer on mixture of experts goes deeper.

Size, speed and cost

For a given design, doubling the parameters roughly doubles the arithmetic needed for each generated token. That has three effects you can feel:

Speed. Larger models respond more slowly unless they run on more hardware. Our article on AI latency covers the other causes of delay.
Price. Serving a large model costs more per answer, which is one reason assistants meter usage. Ask Mio uses points, with a normal chat reply costing 1 point, a long or complex one 3 and a coding answer 5 to 10, as explained in tokens vs points.
Energy and hardware. Bigger models need more electricity and more specialised chips.

Training is far costlier than answering a single question. A common rule of thumb says training compute scales with parameters multiplied by the number of training tokens, which is why only a few organisations train the largest models from scratch.

Open weights: when you can see the parameters

An open-weight model is one whose parameters are published, so anyone can download and run it. A closed model keeps them private and is available only through the provider’s service. Open weights let you check the size, run the model yourself and fine-tune it, while closed models trade that visibility for convenience. The practical differences are covered in our guide to open-weight and closed models.

What the number means when you choose an assistant

For most people, the number of parameters is not the thing to optimise. Three questions matter more:

Is it good at my task? Test with five of your own real prompts, not a leaderboard.
Is it fast and affordable enough? A smaller model is often the right tool for short, simple tasks.
Is my data handled properly? Size says nothing about privacy.

This is why Ask Mio does not ask you to pick a model. Mio routes each request to the model it judges best for the task: faster models for chat, models suited to code for programming, long-context models for documents and image models for design. Paid plans get the stronger models. You still decide what you want, and Mio decides how to deliver it. Our guide to model routing describes the idea.

A worked example: small model or large model?

Imagine two jobs. The first is turning a customer’s e-mail into a three-line summary. The second is reviewing a 40-page contract for conflicting clauses. A small model handles the first quickly and cheaply, and a larger one would add cost and delay for no visible gain. The second job needs long context and careful reasoning, where a stronger and larger model is more likely to catch a subtle conflict, and where you still verify the result yourself.

The same logic applies to your own workflow. Use a quick, lighter option for drafting, rewriting and sorting, and save the heavy option for complex analysis. When an assistant routes for you, as Mio does, you get that split without managing it. If you run models yourself, start small, measure quality on a sample of real tasks and move up only if the small one fails. The large language model article gives further background on how these systems are built and evaluated.

Common misunderstandings about parameters

“Parameters are facts.” They are numbers that encode patterns, not stored sentences.
“Bigger is always smarter.” Data, training and design change the result as much as size.
“A model with more parameters knows more recent events.” Recency depends on the training cut-off, not the size. See knowledge cut-off explained.
“Parameter count equals context length.” They are separate properties.
“Quantized models are broken.” Careful compression can keep quality close to the original at a fraction of the memory.

Frequently Asked Questions

What are AI model parameters in simple terms?

They are the numbers a neural network adjusts while it learns. Each one slightly changes how strongly part of the network responds to its input. Together, billions of them determine how the model turns your text into a reply. They are not stored facts, but patterns spread across the whole network.

What does 70B mean in an AI model name?

It means roughly 70 billion parameters. The B stands for billion. The figure is a rounded size label, useful for estimating memory and cost, but it does not tell you how good the model is at your task or how it was trained, so test it on real prompts.

Is a model with more parameters always better?

No. Larger models usually have more capacity, but training data, training after pre-training, architecture and specialisation matter as much. A smaller, well-trained model can beat a larger, poorly trained one on a given task. Judge models by their results on your own work, not only by size.

How much memory does a model with 7 billion parameters need?

About 14 GB for the weights in 16-bit form, about 7 GB at 8 bits and about 3.5 GB at 4 bits. You need extra memory for the conversation and overhead, so real requirements are higher. The calculation is parameters multiplied by bytes per parameter.

What are active parameters in a mixture-of-experts model?

They are the parameters actually used for each token. A mixture-of-experts model has a large total count but activates only a few specialist sub-networks per token. Total parameters set the memory needed, and active parameters set the computation per token, which is why such models can be large yet fast.

The Bottom Line

AI model parameters are the learned numbers that define a model, and their count is a useful guide to size, memory and cost, but a weak guide to quality on your task. Estimate memory with simple multiplication, remember that data and training matter as much as size, and test on your own prompts. You can try Mio on the free plan without choosing a model, and compare plans on the Ask Mio pricing page.

AI Model Parameters Explained: What the Billions Mean