Learn

The AI glossary.

Plain-English definitions of the words that come up when new AI models launch. Anywhere on trckr, hover over or tap a word with a dotted underline to see its definition.

A

Active parameters: In a mixture-of-experts model, the part that actually works on each word. A model with 74B parameters and 9B active needs the memory of a 74B model but runs about as fast as a 9B one.
Agent: An AI that doesn't just answer but takes steps on its own to finish a task: searching the web, running code or using apps.
AI model: The trained program behind an AI product. ChatGPT is an app; the GPT models inside it do the thinking. One model can power many apps.Labs release new models all the time, and apps switch to them. That's why the same app can suddenly get better, or why one company's model turns up inside another company's product.
API: How developers plug a model into their own apps and pay per use. A model that's "in the API" can be used by programmers, even if it isn't in a chat app yet.

B

Benchmark: A standard test used to compare models, like an exam. Useful, but labs tend to show the tests they do well on, so one score never tells the whole story.

C

Checkpoint: A saved copy of a model at one point in its training. Labs share checkpoints for researchers; they're rarely finished products.
Closed model: A model you can only use through its maker's apps or API: the files that make it work stay private. Most ChatGPT, Claude and Gemini models are closed.
Context window: How much a model can keep in mind at once: your question, any files you share and its own replies. Measured in tokens; 1,000 tokens is about 750 words.

D

Diffusion model: The technique behind most image and video generators: start from random noise and clean it up step by step until a picture appears.
Distillation: Training a small model to imitate a bigger one, to keep much of the quality at a fraction of the cost.

E

Early signal: A hint that a model is coming before it's announced: a name spotted in a lab's systems, an anonymous test model, or code that adds support for it. Usually right, sometimes not.trckr marks these as "Unconfirmed" and only alerts you about them if you switch early signals on. When the official announcement follows, the release is marked as confirmed.
Embedding model: A model that turns text or images into lists of numbers so software can find similar things. It powers search and recommendations; you don't chat with it.

F

Fine-tune: Extra training that adapts an existing model to a narrower job, like answering medical questions or writing in a company's style.
Flagship: A lab's best and usually most expensive model, the one it leads with.
Frontier model: The most capable models of the moment, from the handful of labs at the cutting edge.

G

GPU: The chip that runs AI models. Bigger models need more and larger GPUs, which is why the biggest ones live in data centres.

H

Hallucination: When a model states something false with confidence. Every model does it sometimes; newer ones do it less.
Hugging Face: The website where labs and researchers publish models anyone can download. Think of it as an app store for AI models.

I

Image generator: A model that creates or edits pictures from a written description.
Inference: Running a trained model to get an answer. Training builds the model once; inference happens every time someone uses it.

L

Language model (LLM): A model trained on huge amounts of text to read and write. The engine behind chatbots like ChatGPT, Claude and Gemini.
Latency: How long you wait before a model starts answering.
Leaderboard: A ranking of models, either by test scores or by people voting for the better of two anonymous answers.

M

Mixture of experts (MoE): A design that splits a big model into many "expert" parts and uses only a few for each word. It gives big-model quality at closer to small-model speed.
Model card: The page a lab publishes with a model: what it's for, how it was tested and where it falls short.
Model ID: The exact name developers type to use a model, like gpt-5-mini. New IDs often show up in a lab's systems before the official announcement, which is how trckr spots many releases early.
Multimodal: Works with more than text: a multimodal model can also look at images, listen to audio or watch video.

O

Open source: Strictly, the model's files, code and training recipe are all public and free to reuse. Often used loosely to mean "free to download".
Open weights: The model's files are public, so anyone can download it, run it on their own computer and build on it, usually for free. trckr labels these "Open".Running one yourself keeps your data on your own machine. Small ones run on a laptop; the biggest need servers. Check the licence before using one in a product.
OpenRouter: A service that gives developers one place to use hundreds of models from many labs. New models often appear there early, sometimes anonymously.

P

Parameters: The numbers a model learns during training: a rough measure of its size. Bigger usually means more capable, but slower and pricier to run. 7B means 7 billion.
Post-training: The finishing stage after the main training, where a model learns to follow instructions, use tools, be helpful and avoid harm.
Preview: An early version released for testing. It may change or disappear, and is often limited while it lasts.

Q

Quantisation: Shrinking a model by storing its numbers less precisely, so it fits on smaller computers. A little less accurate, much lighter. FP8, GGUF, AWQ and MLX are common formats.

R

Reasoning model: A model that thinks step by step before answering. Slower, but much better at maths, code and tricky questions.
Reranker: A model that sorts search results by how relevant they are. It works behind the scenes in search tools.
Robotics model (VLA): A vision-language-action model sees through a camera, understands instructions and controls a robot's movements.

S

Snapshot: A dated, frozen version of a model, so developers' apps keep behaving the same even when the main model is updated.
Speech-to-text: Turns spoken audio into written text. Also called transcription or ASR.
State of the art (SOTA): The best result anyone has reached so far on a task.
Stealth model: A model released anonymously under a code name, so its maker can test it with real users before the official launch.

T

Text-to-speech: Turns written text into a natural-sounding voice.
Tokens: The small chunks of text a model reads and writes, about three-quarters of a word each. Prices and limits are counted in tokens.
Tool use: A model's ability to use other software, like a calculator, a web search or a booking system, instead of only writing text.

V

Video generator: A model that creates video clips from a written description or a still image.
Vision-language model: A model that can look at images and talk about them.

W

World model: A model that learns how the physical world behaves, used to simulate scenes for robots, games and video.

Now see them in the wild.

Every release on trckr starts with one sentence in plain English, with the technical details underneath.

Browse the latest releases