inferencia

in English: inference

basicwidespread

Inference is the moment when a trained AI model works: generating an answer, image, or decision from your query or other input.

In plain terms

An AI model has two major phases in its life. The first is training: during this, it learns patterns from countless examples. The second is inference: during this, it uses what it has learned on a new task. When you ask a chatbot a question and it answers, inference is taking place.

It is like the case of a chef. The training years are long and expensive, but only need to be completed once. After that, preparing each ordered lunch is a separate task: small, fast, yet repeated hundreds of times daily. Training is the culinary school; inference is the cooking.

This is why it is discussed so much in the news. Every question, every generated image, every translation is a separate calculation that consumes machine time and electricity. A popular service involves many millions of runs daily, which determines how much AI costs and how fast it responds.

A key difference is that during inference the model typically does not learn. Its knowledge remains the same; it only applies what it previously acquired.

How it works

You type your question or upload an image. The system converts this into numbers, breaking the text into small pieces called tokens.
Numbers pass through the layers of the model. The model calculates based on its internal values set during training, its parameters, and does not change these during the process.
At the end, the model calculates the probability of possible outcomes. In a language model, this means determining which word fragment could follow; in image recognition, for example, whether a cat is in the image. In text, the system does not always choose the most probable word fragment, so slightly different answers can be generated for the same question.
With text, this step is often repeated: the model builds the answer piece by piece (token by token), which is why you see the text gradually appear on the screen.
This mostly runs on special chips in a remote data center, but with smaller models it can even happen on your own phone or laptop.

An everyday example

You photograph a foreign-language menu, and your phone displays it in Hungarian within seconds. At this moment, the translation model performs inference: it learns nothing new, but applies previously acquired knowledge to your image. The same happens when your email client filters spam or a chatbot writes an email for you.

Who is affected

everyoneWhenever you use an AI tool, inference runs in the background. This determines how quickly you get an answer and how much you can use the service for free.

entrepreneur, executiveAI service fees are mostly calculated based on usage, so the cost of inference appears directly on your monthly bill.

developerYou must decide where and on which model inference will run, as this determines speed, cost, and data protection. → Code

A common misconception

MythThe model learns from every conversation, so what I enter today, it will know tomorrow.

In realityDuring inference, the model's internal knowledge does not change; it only applies what it learned previously. Within a conversation, it can use what you wrote earlier because the system receives that text again at every step. Some services have memory functions: in such cases, the system stores notes about you and later provides them to the model, but the model itself does not change from this. The provider may use your conversations for future training, but this is a separate process determined by data processing settings.

Limits and risks

The model only knows what it learned by the end of training; it knows about recent events only if the system connects search or another external source.
The answer is a probabilistic estimate, not a verified fact, so it can be wrong even when sounding confident.
If inference runs in a remote data center, your entered data leaves your device, so check the data processing terms before providing sensitive information.
Every run consumes electricity and machine time, so providers may limit usage, or response times may slow during peak hours.

Similar, but not the same

training – During training, the model's knowledge is formed and evolves; during inference, the finished model only uses this knowledge.
fine-tuning – Fine-tuning is supplementary training that modifies the model, whereas inference does not touch it.
statistical inference – In statistics, inference means drawing conclusions about a whole population from a sample; in AI, it refers to running a trained model.

Related concepts

Broader

machine learning, artificial intelligence

Narrower

on-device AI, batch processing, real-time inference

On the edge

training, modell, token, GPU, latency

A brief history

The word originates from the Latin verb "inferre" (to bring in, to infer) and has long been the technical term for inference in logic and then statistics. In AI, the early expert systems' "inference engine" used it, and with the spread of machine learning, the term began to refer to running a trained model. After the launch of ChatGPT in 2022, the word became widely known in news about AI costs, chips, and energy demand.

Where things stand

As of late 2026, inference is a daily mass service: chatbots, search engines, phones, and office programs perform it continuously. The so-called reasoning models compute for longer before answering, so each response requires more computation. Industry attention is increasingly focused on how to run models more cheaply, faster, and with less energy.

Status: 2026 November

In the news

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

China approves overseas use of DeepSeek and MiniMax models as DeepSeek open-sources its inference engineOct 19, 2026 · Agent Age
Lovable offers EU inference to control where AI requests are processedOct 12, 2026 · Code
OpenAI unveils Jalapeño inference chip, promising faster AI responsesAug 25, 2026 · Agent Age

Also known as

inferencia · inference · inference · AI-inferencia · MI-inferencia · model inference · model inference · AI inference

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