AI model glossary: 37 terms in plain words

Model pages are full of shorthand: 9B, A3B, native multimodal, per MTok, context window. This glossary explains the terms you’ll meet on spec sheets, ours included, in plain words, with links to longer explanations where we have them.

Size and architecture

Parameter

One learned number inside a neural network: a weight in one of its matrices. A model’s size is its parameter count, so 9B means nine billion. More parameters mean more capacity to store knowledge, and more memory and compute to run. See Why 9B?

Dense model

A model that uses every one of its parameters for every token it processes. Its total and active parameter counts are the same.

Mixture of experts (MoE)

An architecture that splits parts of the network into many “experts” and uses a router to send each token through only a few of them. It computes like a much smaller model but has to keep all of its experts in memory.

Active parameters

In a mixture-of-experts model, the parameters actually used for each token. In a name like 30B-A3B, 30B is the total and A3B means about 3 billion are active.

Transformer

The neural-network architecture behind almost every current language model, introduced in 2017 in “Attention Is All You Need”. Its key mechanism, attention, lets each token weigh the other tokens in its context.

Quantization

Storing a model’s weights with fewer bits, such as 8 or 4 instead of 16, to cut memory and speed up inference, at some cost in accuracy.

Distillation

Training a smaller “student” model to imitate a larger “teacher”, so that it keeps much of the teacher’s ability at a fraction of the size.

Tokens and context

Token

The unit a language model reads and writes: a word, part of a word, a punctuation mark or a piece of code. In English a token averages about four characters, or three quarters of a word. See LLM API pricing explained.

Tokenizer

The component that splits text into tokens from a fixed vocabulary. Models use different tokenizers, so the same text can come to a different number of tokens.

Context window

The maximum number of tokens a model can take into account at once: the prompt, the conversation so far and its own reply.

KV cache

The stored attention state (keys and values) for the tokens a model has already processed, kept so that it doesn’t recompute them for every new token. It grows with context length and takes up GPU memory.

Input and output tokens

Input tokens are everything sent to the model; output tokens are what it generates. APIs price them separately, and output usually costs more.

Price per million tokens (per MTok)

The usual unit of API pricing. $0.04 per MTok of input means four cents for every million tokens sent. See KTwo’s pricing.

Reasoning tokens

Tokens some models generate to work through a problem before they answer. They are often hidden from the user but billed as output.

Training

Pre-training

The first and largest stage of training, in which a model learns from a huge body of text (and, for multimodal models, other data), usually by predicting the next token.

Post-training

Everything after pre-training that turns a raw model into a useful assistant, such as instruction tuning, preference tuning and safety training.

Fine-tuning

Further training of an existing model on a smaller, targeted dataset, to change its behavior or teach it a domain.

Instruction tuning

Fine-tuning on examples of instructions and good responses, so that a model follows requests instead of merely continuing the text.

Reinforcement learning from human feedback (RLHF)

Training a model with a reward signal derived from people’s preferences between its answers. Related methods learn from AI-written preferences, or optimize on preferences directly.

Knowledge cutoff

The date after which a model’s training data contains no information. A frozen model knows nothing that happened later.

Self-training

Learning in which a model’s own outputs, filtered for quality, become its training data. KTwo Crust 1 Beta trains itself natively. See What is a self-training AI model?

Continual learning

Training a model on new data over time without starting again from scratch, while keeping what it already knows.

Catastrophic forgetting

The tendency of a neural network to lose earlier skills when it is trained on new data: the central problem of continual learning.

Model collapse

The degradation that comes from training models repeatedly on model-generated data without enough filtering: rare knowledge vanishes and output turns repetitive.

Synthetic data

Training data generated by models rather than written or collected by people.

Modalities

Multimodal

Able to take in, and sometimes produce, more than one kind of data, such as text, images, audio or video.

Natively multimodal

Trained on several modalities together from the start, as one model, instead of having a vision or audio component attached to a finished language model. KTwo Crust 1 Beta is natively multimodal. See Natively multimodal models, explained.

Vision encoder

A network that turns an image into a sequence of feature vectors that a language model can read. Add-on multimodal systems connect one to a language model.

Early and late fusion

Early fusion combines modalities from the start, in one model; late fusion processes them separately and combines the results afterwards.

Vision-language model (VLM)

A model that takes images and text as input and answers in text. Nemotron 3.5 Super VL, free in LE PEAK, is one.

Using models

System prompt

Instructions sent with every request that set a model’s role, rules and tone. They count as input tokens on every call.

Retrieval-augmented generation (RAG)

Fetching relevant documents from a search index and adding them to the prompt, so that a model can answer from them. The model’s weights don’t change.

Hallucination

A fluent but false statement from a model, such as an invented fact, quote or citation.

Inference

Running a trained model to produce output, as opposed to training it.

Latency and throughput

Latency is how long a single answer takes, often measured as the time to the first token; throughput is how many tokens a system produces per second across all its users.

Open-weight model

A model whose trained weights are published for anyone to download and run. gpt-oss 120B, which you can use in LE PEAK, is one.

API ID

The exact string used to request a model through an API. KTwo Crust 1 Beta’s is ktwo-crust-1-beta; see All models.