weight
A weight is a number in an artificial neural network that indicates how much a connection matters; the model's knowledge resides in these numbers.
In plain terms
Imagine a huge mixing console with billions of sliders. Each one controls how strongly a given signal passes to the next station. These sliders are the weights, and the language model is essentially the sum of their settings.
Weights are not set by humans. During training, the model practices on vast amounts of text and repeatedly adjusts the sliders slightly based on its errors. By the end, a configuration forms that can produce meaningful sentences, translations, or program code.
A weight is not a database and not a collection of rules. From a single weight, one cannot read that 'Budapest is the capital of Hungary': knowledge is present in a distributed manner, in the combined pattern of billions of numbers. This is why it is difficult to explain why the model gave the answer it did.
How it works
- A neural network consists of small computing units (neurons) connected by links; each link has an associated weight.
- When you type a question, your text is converted into numbers and passes through the network: at each step, incoming values are multiplied by weights and then summed.
- A large positive weight amplifies the signal, one near zero almost ignores it, and a negative one works against it.
- During training, the model guesses the next word, its guess is compared with the correct solution, and based on the error, every weight is shifted slightly in the right direction. This is repeated countless times while the model sees text amounting to billions of words.
- At the end of training, the weights are fixed and saved to a file. When you converse with the model, these no longer change; the model only uses them.
An everyday example
You read in the news that a company released an 'open-weight' model that anyone can download. This means they published the file containing the trained numbers, which can range from a few gigabytes to several hundred gigabytes depending on the model size. Whoever downloads it can run the model on sufficiently powerful local hardware or servers, ensuring the entered data does not reach the model creator.
Who is affected
A common misconception
Limits and risks
- Weights are frozen at the end of training, so the model cannot know on its own about events that have happened since.
- At present, only part of what the model knows, and why, can be read out of its weights, so its errors cannot be fixed by changing a single number.
- More parameters do not automatically mean a better model: the quality of training data and the training method matter just as much.
- An 'open-weights' model is not necessarily open-source: the license may restrict usage, and the training data or training code is often not public.
Similar, but not the same
- parameter – Parameter is the broader term: it includes weights as well as bias values, although in everyday usage the two are often interchanged.
- Hyperparameter – The developer sets the hyperparameter before training (for example, the learning rate), while the model itself learns the weight.
- token – A token is a piece of text processed by the model, while the weight is the model's internal setting with which it processes the tokens.
- attention weight – Despite its name, the attention weight is not a learned, fixed number: the model calculates on the fly for every text how much to focus on each word. The model weight, however, does not change after training.
Related concepts
training, fine-tuning, backpropagation, quantization, Hyperparameter, bias (offset value), attention weight
A brief history
The idea of weighted connections was born from the model of brain synapses: Frank Rosenblatt's perceptron described in 1958 already learned with adjustable weights. The training method still used today, backpropagation, was made widely known by a 1986 study by Rumelhart, Hinton, and Williams.
Where things stand
As of late 2026, every significant language model is built on weights. The largest publicly known models use several hundred billion, or even around a trillion, parameters, while companies usually do not disclose the exact size of closed models. In the news, you mostly encounter the term in connection with "open-weight" models and the secure safeguarding of weights.
In the news
- Hungarian user Vera Mérő sought weight-loss help and received life-threatening advice from an AI chatbotNov 3, 2026 · Health
- 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
- Skilled trades worth their weight in gold, Hungarian managers squeezed out: new fault lines in the domestic labour marketOct 26, 2026 · Work & HR
- Skilled workers are increasingly valuable in Hungarian industry, says Job Impulse headOct 22, 2026 · Ipar
- Mistral unveils 1-trillion-parameter Large 4, plans open-weights releaseOct 6, 2026 · Agent Age
- Justice Department backs OpenAI's fair-use defense in copyright case; news publishers objectSep 28, 2026 · Jog
- South Korea's 2027 budget plan sets aside 1.26 trillion won for regional and essential healthcare, with more medical AISep 1, 2026 · Health
Also known as
weight · weights · model weights · weight · weights · model weights · synaptic weight · open weights
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.