AI Academy · Additional advanced reading · Educational pilot

Eight AI Labels: Model Families and Methods

Not every AI model generates text. But these eight abbreviations do not describe eight competing architectures. Learn what each label means, where they overlap and how to choose technology for a real task.

Original teaching material informed by a supplied infographic. Simplifications have been corrected; independent educational review is pending. This supplementary guide is not a complete course.

AI Foundations overview

Five different ways to describe AI

The same system can fit several labels. A compact language model may use mixture-of-experts layers and tools, while a vision-language model may work with a separate segmentation model. Classify by the dimension being described, not by acronym alone.

Language and vision capability

LLM, VLM

Scale or deployment

SLM

Architecture

MoE

Training or generation method

MLM, LCM

Action systems or named model family

LAM (informal), SAM

Explore the eight labels

LLM Large Language Model

Capability: language

Processes language; many modern generative LLMs predict a next token repeatedly.

When useful
Drafting explanations, summarizing text, software assistance.
Watch out for
Fluent text is not proof of truth. A model does not automatically have tools or current information.
Overlap
An LLM can also be small or use MoE.

LCM Latent Consistency Model

Generation method: image sampling

Learns a consistency-based mapping that can generate images in few sampling steps within supported diffusion-family systems.

When useful
Faster image generation where a compatible system is available.
Watch out for
LCM pipelines do not inherently require SONAR embeddings, sentence segmentation or quantization. Output quality must be tested.
Overlap
It is related to diffusion-based image generation.

LAM Large Action Model

Informal label: action-capable systems

Used inconsistently for models or applications that interpret requests and plan or perform software actions.

When useful
Assisted scheduling, form filling and workflow execution with authorization.
Watch out for
There is no one universally standardized LAM architecture. The surrounding application must enforce permissions and confirmations.
Overlap
Many so-called LAM systems use an LLM or VLM plus tools.

MoE Mixture of Experts

Architecture

Routes representations to selected expert modules, often token by token, then combines their outputs.

When useful
Increasing model capacity while activating only some expert components.
Watch out for
Sparse experts do not automatically make the whole application faster or cheaper; communication and memory costs matter.
Overlap
An LLM itself can have an MoE architecture.

VLM Vision-Language Model

Capability: vision plus language

Combines visual inputs with language capabilities, using visual encoders or tokens and multimodal fusion.

When useful
Image questions, diagram description and document interpretation.
Watch out for
Small text, unusual diagrams and subtle details can be misread. Check important results.
Overlap
A VLM may also be an LLM and may use MoE.

SLM Small Language Model

Scale and deployment

A relatively compact language model; model size is a spectrum, not a fixed universal parameter limit.

When useful
Potentially useful on laptops, mobiles or embedded devices with adequate resources.
Watch out for
Smaller does not always mean acceptable accuracy, faster performance or lower energy on a specific device.
Overlap
An SLM remains a language model and can use several architectures.

MLM Masked Language Modeling

Training objective

Hides or replaces selected tokens and trains a model to predict them from available context, as in BERT training.

When useful
Learning representations for classification and search after appropriate adaptation.
Watch out for
A masked-language training objective is not the same as autoregressive open-ended text generation.
Overlap
MLM is an objective used to train language models, not a rival species to LLM.

SAM Segment Anything Model

Named segmentation model family

Predicts pixel-level masks from images and supported prompts using a promptable segmentation system.

When useful
Object masks, annotation support and image editing.
Watch out for
A mask is not proof of semantic understanding, correct object identity or suitability for medical or autonomous safety-critical use.
Overlap
SAM refers to a particular family and task, not a model-size label.

Compare two practical systems

A compact language assistant: It can be a language model (LLM), be relatively small (SLM), and use an MoE architecture. These are not mutually exclusive options.

A visual workflow assistant: It could use a VLM to interpret an image, SAM for proposed image regions, and a host application with explicit user approval before taking consequential actions.

Try identifying the right concept

Predict the most relevant label or approach for each scenario, then open the explanation.

1. An offline assistant on a constrained device

An SLM might be worth testing. Measure accuracy, memory, latency and power on the target hardware.

2. An image editing tool needs object masks

SAM or another segmentation model family could be evaluated; validate the masks on representative images.

3. A chatbot routes token computations to selected experts

That describes MoE, which is an architecture that may be part of an LLM.

4. A training task hides words and predicts them using context

That describes MLM, the masked-language training objective.

5. Software books a meeting after a user authorizes it

This is an action-capable agent workflow, sometimes marketed as LAM; safe execution requires host-side controls.

6. A model describes a diagram that a learner uploaded

A VLM may be appropriate, with checks for visual errors.

7. An image generator needs few inference steps

A compatible LCM-style method may be relevant; speed and quality must be measured.

8. A system drafts text explanations

A generative LLM may fit, but the explanations still need verification.

Engineering challenge

A community learning center wants an offline assistant to help interpret diagrams and organize donated books. Sketch two possible solutions. Decide which tasks can be handled by simple rules, which need vision, which need language, and what should require a human decision.

Describe your inputs, device limitations, permissions, baseline method, testing plan, latency and failure recovery. Compare candidate systems using evidence, not the biggest model name.

Further reading

Research and technical references for several of the terms:

Model output is not a medical diagnosis, authorization to operate machinery or evidence of safety-critical reliability. Evaluate each system in its actual use context.