Learn AI /Large Language Models /Parameter
Parameter
- Starter
- Large Language Models
- Policy officers
- Managers & front offices
Also called: parameters, parameter count.
Back to the milestone you came from
In five seconds
What it is
A parameter is a learned numerical setting inside a model that helps shape which token it is likely to produce next.
In plain words
Parameters are the many learned numbers inside a model that help it guess what text should come next.
An everyday example
Imagine a phone keyboard that adjusts many hidden numerical preferences as it learns from repeated text and corrections. Later, when you type “See you in the”, those settings work together with the words already entered to rank possible next words. One suggestion may be much more likely than the others, but no single setting contains the whole phrase or decides the answer alone. A language model uses vastly more parameters and more complex calculations, but the basic idea is similar.
A student can explain reasons and use broader world knowledge across situations. Model parameters are only numerical values adjusted during training. They do not work like human study habits, memory, or conscious strategies.
Remember thisParameters are learned numerical values that jointly shape next-token probabilities. Their number indicates model scale, not whether the model is accurate, suitable or safe.
Understand it
Understand it
Why it matters
If you hear that a model has billions of parameters, that tells you something about its size and cost, but not whether it is safe, accurate, or suitable for your task. This helps leaders ask better questions when comparing tools, budgets, speed, and risk.
How parameter changes shape next-token choices
After using this explainer, a reader should be able to say that parameters are learned numbers inside a model, and changing them changes the probabilities of possible next tokens.
A compact simulation shows a short prompt, three candidate next tokens, and sliders representing groups of learned parameters. As the reader moves sliders, the token probability bars change and one token becomes the sampled output. The explainer focuses on influence, not on exact maths.
Readers can move several sliders representing groups of parameters, reset the simulation, and press a button to sample a next token from the current probabilities.
Describe this without the diagram
Screen-reader mode presents the same prompt in text, then lists candidate next tokens with current probabilities. It explains that each slider represents a group of learned parameters. When a slider value changes, the new probabilities are announced in words, for example: 'Token A now 55 percent, Token B 30 percent, Token C 15 percent.' A button then reports which token was sampled. The narration makes clear that parameters do not appear in the final text directly; they influence probability calculations inside the model.
With reduced motion: For readers who prefer less motion, the same prompt, sliders, percentages, and selected token appear without animated bar movement or transitions. Probability changes update instantly as static values.
In a large language model, text goes in as tokens, which are small pieces of text such as words, parts of words, or punctuation. Inside the model are many numerical settings. These learned settings are called parameters.
During training, the model is shown large amounts of text and tries to predict the next token. When it predicts badly, training adjusts many parameters slightly. Over time, those adjustments make some patterns more likely and others less likely.
Later, when you enter a prompt, the model uses its parameters together with the current context to calculate probabilities for possible next tokens. The output is a stream of chosen tokens, which becomes the reply you see.
People often confuse parameters with tokens or with hyperparameters. Tokens are pieces of input and output text. Hyperparameters are training choices set by engineers, such as learning rate or batch size. Parameters are the learned numbers produced by training.
How “Predictive text learning your phrasing” maps onto the real thing
| In the story | In the AI system |
|---|---|
| many small study-habit adjustments | many learned parameters |
| teacher reviews mistakes after each practice paper | training compares predictions with correct next tokens and updates parameters |
| new practice paper | a new prompt and context |
| which answer she writes next | which next token the model is likely to produce |
A worked example
A small demonstration model is trained on public administrative sentences, including examples ending with the phrase “close of business”. It is then prompted: “Please send comments by close of”.
Starting from: Prompt: “Please send comments by close of”
- The model converts the prompt into tokens representing the words and punctuation.
- During training, whenever its predicted next token differed from the token in the training text, optimisation adjusted many parameters slightly to reduce that prediction error.
- Across many examples, those adjustments made some continuations more probable in particular contexts. No single parameter stores the phrase or controls the answer by itself.
- During inference, the learned parameter values are used with the prompt to calculate a probability distribution over possible next tokens, such as “business”, “the” or “day”.
- A decoding method selects one token, adds it to the context and repeats the calculation for the following token.
Result: The model is likely to produce “business” because its learned parameters give that token a relatively high probability in this context. The parameters shape the distribution, but do not make the output certain.
In a United Nations workplace
At a duty station working in three languages, a humanitarian affairs officer is sending a side-event concept note for clearance before close of business. During a vendor demonstration, the officer asks whether the model advertised as having more parameters will necessarily produce the better first draft in all three languages.
Parameters jointly shape the probabilities assigned to possible next tokens. A larger parameter count means more trainable numerical values, but does not by itself establish accuracy, fluency across languages, suitability for the task or acceptable computing cost. Those claims require testing on representative material.
Before you try thisParameter count says nothing about how a tool handles data. Do not enter uncleared positions, personal data or security-sensitive details into a service that has not been approved for them. Use only approved tools, test outputs in each working language and retain human review before clearance.
Where you may hear it
- A vendor demonstration during a procurement unit’s technical evaluation
- An ICT governance board reviewing hosting and computing requirements
- A front-office briefing before approval of an AI drafting pilot
- A technical review of model performance before deployment
Avoid confusion
Avoid confusion
Often confused with
Weight
A weight is a learned number that scales how strongly one input signal affects the next calculation in a neural network.
Hyperparameter
A hyperparameter is a training setting chosen before or during learning that shapes how a model is fitted, rather than a value the model learns from data.
Token
A token is one unit from a model’s vocabulary used to represent input or output, often smaller than a word.
Side-by-side
Common misconceptions
Common misconception: More parameters always mean a better model.
Correction: A larger model can capture more patterns, but quality also depends on training data, training method, evaluation, and how the model is used. Bigger can also mean slower and more expensive.
Common misconception: Parameters are the same as the words stored in the model.
Correction: Parameters are numerical settings, not a visible list of stored sentences. They influence probabilities for outputs rather than acting like a simple database of exact text.
Common misconception: Parameter count tells you whether a model is trustworthy.
Correction: Trustworthiness depends on many factors, including testing, safeguards, domain fit, and human oversight. Parameter count alone says little about factual reliability or safety.
Quick check
Which statement best describes parameters in a large language model?
Parameters are learned numbers inside the model. Tokens are pieces of text, manual rules are a different approach entirely, and context window size is a separate design feature about how much text the model can use at once.
Go deeper
Go deeper
The simple version
Parameters are the many learned numbers inside a model that help it guess what text should come next.
The technical version
Parameters are the trainable numerical values of a neural network. During training, optimisation updates these values to reduce prediction error, and during inference they are used to compute probability distributions over next tokens.
Reviewed external resources
EncyclopaediaParameter (opens in a new tab)Wikipedia
This Wikipedia page gives a broad general definition of parameter across fields, which suits readers who want a neutral starting point.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This Hugging Face course material helps readers place parameters inside how large language models are trained and used, and suits beginners who want guided learning.
Deeperlink checked 2026-07-31
VideoLarge language models, explained visually (opens in a new tab)3Blue1Brown
This 3Blue1Brown video offers an intuitive visual explanation of how model internals such as parameters relate to behaviour, and suits visual learners.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Parameter (opens in a new tab)YouTube search
This YouTube search results page helps readers find multiple video explanations of parameter from different teachers, which suits people comparing styles before choosing one.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Parameter (opens in a new tab)Reddit search
This Reddit search results page helps readers see common questions and informal explanations about parameter, which suits readers who want to spot where people get confused.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Parameters are learned numerical values that jointly shape next-token probabilities. Their number indicates model scale, not whether the model is accurate, suitable or safe.