model weights

basicwidespread

Model weights are the numbers in which an AI model stores the knowledge acquired during training; these determine what it answers to a question.

In plain terms

An AI model internally does not store sentences or rules, but vast numbers. These are weights: each indicates how strongly one small computing unit of the model should affect another. Today's large language models contain several billion such numbers; the largest have several hundred billion, or even over a trillion.

Imagine a massive mixing console with billions of sliders. During training, the machine nudges each one slightly, over and over, until the result sounds good. The final setting is the model: if you copy the slider positions, you have copied the model.

That is why the weight file is so valuable. Training can take months and cost a fortune, yet the result is a single file package that fits on a hard drive. Whoever obtains it can run the model without paying the cost of training.

Weights are not the same as the source code or the training data. The code is the skeleton, the data is the curriculum, and the weights are what remains after learning.

How it works

At startup, the weights are random numbers, so the model initially gives nonsensical answers.
During training, the model receives examples, guesses the correct answer, and the system measures how far off it was.
Based on an error, a mathematical procedure adjusts every weight by a hair's breadth in the direction that reduces the mistake. This step is repeated countless times, even millions of times, during training.
At the end of training, weights are fixed and saved to a file. During use, they no longer change: your query passes through them, and the answer is assembled from the calculations.
The trained weights can later be further adapted: fine-tuning tailors them to a specific field, while compression makes them runnable on smaller machines, at the cost of some accuracy.

An everyday example

You download a freely available language model to your laptop to chat without internet. The multi-gigabyte file, which takes minutes to download, consists almost entirely of the model's weights. The runtime program is negligible in size.

Who is affected

developerThey download, run, fine-tune and compress the weights, and must also make sure they come from a trustworthy source. → Code

entrepreneur, executiveA decision must be made whether to rent a provider's closed model or run an open-weight model on the company's own machines, where data remains in-house.

lawyer, citizenThe release, protection, and export of weights have become a regulatory issue, because published weights cannot be withdrawn. → Law

everyoneEvery response you get from an AI chatbot consists of these numbers, so the model's knowledge and biases are embedded within them.

A common misconception

MythThe weights literally contain the texts used for training, like in a database, and the model retrieves the answer from there.

In realityWeights compress patterns and relationships into numbers; they do not store documents. The model can reproduce some frequently repeated text segments verbatim, but this is a side effect, not the basis of its operation.

Limits and risks

The weights do not reveal what the model knows or why it responds as it does: researchers only partially understand what each number signifies.
Weights capture the moment training ends, so the model cannot learn about events that occur later on its own.
Released weights cannot be recalled, and built-in safety limits can be removed from them via fine-tuning.
Downloading weight files from unknown sources is risky: some older file formats may contain malicious code, and the model itself could be intentionally manipulated.

Similar, but not the same

parameters – Parameter is the broader collective term for everything the model has learned; weights make up the vast majority of these, which is why everyday usage treats the two as synonyms.
training data – Training data is the text, image, or audio from which the model learns, while weights preserve the result of learning, not the training material itself.
open-source model – Publishing weights does not mean open source: the training data and training code can still remain secret, and usage may be restricted by license.

Related concepts

Broader

neural network, machine learning, MI-modell

Narrower

open-weight model, quantization, LoRA-adapter, checkpoint

On the edge

parameters, training, fine-tuning, training data, inference

A brief history

The idea of adjustable weights dates back to Frank Rosenblatt's perceptron described in 1958, which already learned by modifying the strength of connections. The efficient adjustment of weights in multilayer networks was made possible by backpropagation, which was widely popularized by a 1986 study by Rumelhart, Hinton, and Williams.

Where things stand

At the end of 2026, every major AI model, including language, image and audio models, is built on weights, and a large number of freely downloadable open-weight models are available. However, the weights of the flagship models of the largest companies are strictly guarded trade secrets. A security-policy debate is under way about whether these should be released and how they should be protected.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

DeepSeek's agent platform reaches 1.54 million daily users as Chinese open models spreadNov 2, 2026 · Agent Age
Mistral Large 4 is the best open model outside China, with weights released publiclyOct 29, 2026 · Agent Age
Mistral unveils 1-trillion-parameter Large 4, plans open-weights releaseOct 6, 2026 · Agent Age

Also known as

model weights · weights · model weights · neural network weights · model weights · weights · open weights · open weights

Sources

Rumelhart, Hinton, Williams: Learning representations by back-propagating errors (Nature, 1986)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.