Model icons

The 59 model icons in AI & LLM, described

Base modelbase-model

A base model — the pretrained foundation before anybody tuned it.

Cascadecascade

Model cores stepping down a diagonal — a cascade that escalates from cheap models to strong ones.

Cheap modelcheap-model

A small model core beside a coin — the inexpensive model you route easy requests to.

Decoder-onlydecoder-only

A model core with an arrow leaving it — a decoder-only model that only generates.

Depthdepth

Depth — how many layers a network stacks, how deep the model goes.

Diffusiondiffusion

Diffusion — noise resolved step by step into a picture by a generative model.

Diffusion modeldiffusion-model

The model frame filled with scattered noise — a diffusion model that denoises its way to an image.

Draft modeldraft-model

A small model core beside a large one — the draft model that proposes tokens for the big model to verify.

Edge modeledge-model

A model core tucked into the corner of a frame — a model that runs at the edge, near the user.

Encoder-decoderencoder-decoder

Encoder-decoder — one side reads the input and the other writes the output, a seq2seq pair.

Encoder-onlyencoder-only

An arrow entering a model core — an encoder-only model that reads and represents.

Ensembleensemble

Three model cores arranged together — an ensemble whose answers are combined.

Fallback modelfallback-model

A model core with a smaller one hanging beneath it — the fallback used when the first is unavailable.

Frontier modelfrontier-model

A model core beneath a mountain peak — the most capable model at the frontier.

Instruct modelinstruct-model

An instruct model — tuned to follow what you ask, aligned for chat.

Latentlatent

Two triangles meeting at a narrow waist — the compressed latent space between encoder and decoder.

LLMllm

A large language model — the generative AI system behind chat, completion and reasoning.

Modelmodel

A trained machine learning model, the network of weights that turns an input into an output.

Model addmodel-add

A model frame with a plus beside it — register another model in the catalogue.

Model alertmodel-alert

A model frame with an alert mark beside it — something is wrong with this model and needs attention.

Model aliasmodel-alias

A model frame with a label tag inside — an alias like latest that points at a version.

Model archivemodel-archive

A model core above an open storage tray — a retired model archived for the record.

Model bookmarkmodel-bookmark

A bookmark beside a model core — a model saved to come back to.

Model cardmodel-card

A model card — what this model is, how it was trained and where it is safe to use, on one page.

Model checkmodel-check

A model frame with a check beside it — the model passed evaluation, verified and ready to serve.

Model deploymodel-deploy

Model deploy — put a trained model where it will be served and used.

Model downloadmodel-download

Model download — fetch the weights and pull a model onto this machine.

Model failmodel-fail

A model core with an X inside — a model call that failed and returned an error.

Model familymodel-family

A model family — one base model and its several descendants in a lineage tree.

Model fastmodel-fast

A model core beside a lightning bolt — the low-latency mode of a model.

Model forkmodel-fork

A model fork — take the weights and go your own way with a diverging fine-tune.

Model guardmodel-guard

A model core beside a shield — the guardrails wrapped around a model.

Model haltmodel-halt

A model core beside a stop square — halt a model mid-generation.

Model heartmodel-heart

A model core beside a heart — a favourite model you keep coming back to.

Model idlemodel-idle

A model core beside a Z — a model loaded into memory but doing nothing right now.

Model keymodel-key

A model core beside a key — the API key that unlocks a model.

Model licensemodel-license

A model license — the terms that say what you may do with the weights.

Model lockmodel-lock

A model core beside a padlock — a model locked to approved use.

Model offmodel-off

A model frame struck through — the model is disabled and no longer serving requests.

Model registry entrymodel-registry-entry

A model registry entry — one model on the shelf, versioned and picked out.

Model scopemodel-scope

A model core held between brackets — the scope a model is allowed to work in.

Model searchmodel-search

A model core beside a magnifying glass — search for a model in a catalogue.

Model sizemodel-size

Model size — how big the model actually is in parameters and footprint.

Model swapmodel-swap

Model swap — put a different model behind the same call without changing the caller.

Model tagmodel-tag

A model core beside a label — a tag that groups models by family or purpose.

Model weightsmodel-weights

Model weights — the parameters that are the model, saved as a checkpoint file.

Multi-headmulti-head

Three heads rising from one model core — multi-head attention.

Pin modelpin-model

A model core beside a pin — pin a deployment to one exact model version.

Quantizequantize-4bit

Quantize — the smooth line told to pick a step, weights compressed to fewer bits.

Quantized modelquantized-model

A model frame with a staircase inside — a model quantized to fewer bits so it runs smaller and faster.

Reasoning modelreasoning-model

A reasoning model — one that thinks through a problem in deliberate steps before answering.

Region modelregion-model

A location pin beside a model core — the model served in a given region.

Small language modelsmall-language-model

A small language model — compact enough to run on a device or at the edge.

Sparsitysparsity

Sparsity — a matrix that is mostly zeros, on purpose, so it computes faster.

Tiny modeltiny-model

The model frame, drawn small — a compact model that fits on a phone or at the edge.

Transformertransformer

A transformer — the attention-based architecture under all modern language models.

Transformer blocktransformer-block

A transformer block — attention, then the thinking layer, stacked again and again.

Vision encodervision-encoder

A picture beside a model core — the vision encoder that turns pixels into tokens.

Weight pruneweight-prune

Weight prune — the connections nobody will miss, cut to make the model sparse.

In code, each is one import — import { BaseModel } from "@iconmind/react/icons/base-model" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag