Inference

Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary

The phase where a trained model is put to work: it applies its learned parameters to new data and returns predictions, one request at a time.

What is Inference?

Inference is the phase in which a trained machine learning model is put to work: it takes new, unseen data and applies its learned parameters to produce a prediction, classification, or recommendation.

In statistics, “inference” means drawing conclusions about a population. In machine learning it means something narrower: running a finished model on live input. When engineers talk about inference cost or inference latency, this serving phase is what they mean.

How Inference Works

At inference time the model’s parameters are frozen. Input arrives, flows forward through the model’s computations, and comes out as an output: a label, a score, a sentence. No learning happens; the model gives the same answer to the same input every time.

Because inference runs on every single request, its cost compounds. A model queried a million times a day pays its inference bill a million times, which is why teams compress models with quantization or prune them for sparsity before serving them at scale or on edge devices.

Inference vs Model Training

Inference is when the model answers; model training is when it learns. The two phases have opposite economics: training is a large periodic compute bill, while inference is a small cost paid on every request for the life of the system.

InferenceModel training
What happensParameters are frozen; the model only produces outputsParameters change with every batch of examples
When it runsEvery time the deployed model receives a requestBefore deployment, and again for retraining
Compute profileSingle requests, where milliseconds of latency matterLarge batches on GPU clusters, for hours or weeks
Typical failureLatency spikes, data drift, serving a stale modelOverfitting, data leakage, mislabeled data

Example of Inference

A bank runs a fraud model trained months ago on historical transactions. You tap your card at a store, and the transaction details – amount, merchant, location, time – become the model’s input.

The model applies its frozen parameters and returns a fraud score in a few milliseconds. Above a threshold, the transaction is blocked; below it, the payment clears before the terminal finishes beeping.

That single scoring pass is inference. The model learned nothing from your transaction; it will only get smarter when the bank retrains it on fresh data.

Related AI terms: Model Training · Quantization · Deploy · Edge Computing

Did you like the Inference gist?

Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.

Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About

Read the Governor's Letter

Stay ahead with Governor's Letter, the newsletter delivering expert insights, AI updates, and curated knowledge directly to your inbox.

By subscribing to the Governor's Letter, you consent to receive emails from AI Guv.
We respect your privacy - read our Privacy Policy to learn how we protect your information.

Browse All AI Terms A–Z

Every term in the dictionary, in alphabetical order. Jump to a letter or scroll the full list.

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

X

Y

Z