Understanding AI: Key terms and what they mean in practice
Artificial Intelligence
5 February 2026
Today, professionals must stay current with both developments in their fields and advancements in technology, which now play a central role in most workplaces.
The rapid evolution of AI is transforming how work is done, increasing the need for professionals to understand new terms and concepts. Learning this language is essential to using AI tools effectively.
This LLM and AI glossary covers essential terms that are becoming standard in professional settings. Understanding them is key to staying current in today’s AI-driven workplace.
AI as a broad concept
AI is often viewed as a single concept, but in reality, it consists of multiple components working together. This glossary defines key terms used in AI and LLMs, many of which are also common in the broader IT industry.
Language model
A language model is a type of AI designed to understand human language, and in some cases generate text that resembles human writing.
Language models use machine learning to analyze large volumes of text and learn how words relate to one another in different contexts. Simply put, a model predicts the most suitable next word in a sentence using the context provided by the surrounding text.
This means they can be used in tasks such as speech recognition, machine translation, natural language generation (generating more human-like text) and more.
Language models can vary in size and capability. The so-called LLMs and SLMs are part of this family. The main difference between them is their parameter count, which dictates their overall complexity and resource requirements.
Model parameters
Model parameters are the numerical values a machine learning model learns during training to map input data to outputs, like predictions or generated text.
During training, the machine learning algorithm adjusts these parameters so the AI model’s outputs increasingly match the expected results. Models with more parameters tend to capture more complex patterns—such as Large Language Models (LLMs), which can have billions of parameters.
Smaller models with fewer parameters are usually simpler and require less computational power, though they may capture fewer complex patterns in the data.
Large Language Model (LLM)
A Large Language Model (LLM) is an advanced type of language model, which is trained on extremely large and diverse datasets, often drawn from many different sources and domains.
It uses billions of parameters to recognize patterns and relationships in text, and based on their size, LLMs can work across topics and tasks without being retrained for each one. LLMs can be used to generate text, analyze and structure unstructured data, extract insights, classify information, support automation, and assist in decision-making processes.
Small Language Model (SLM)
A Small Language Model is a language model designed to be more lightweight and efficient than a large language model. and their use requires lower computational resources, making it suitable for fine-tuning and, therefore, specific tasks, in resource-constrained environments, or use cases where full-scale general-purpose models are unnecessary.
Open-source model
A model whose code, architecture, trained model, and sometimes training data are publicly available under a permissive license.
This transparency allows developers, researchers, and organizations to inspect the model’s structure, understand how it was trained, reproduce its behavior and adapt it for new tasks and applications.
These models support transparency, reproducibility, and collaborative development. However, they can come with challenges such as limited support, potential misuse, and security vulnerabilities.
Open-weights model
An open-weights model is a pretrained AI system whose learned parameters are publicly shared under a license that allows reuse. Unlike fully open-source AI, open weights do not include the full training code or datasets, so developers cannot fully reproduce the training process.
However, they can use it for making predictions on new data, adapting it for specific applications, or for building tools that need its capabilities without retraining from scratch.
Deployment
Deployment is the process of putting a trained AI model into a production environment so it can be used in real-world applications. It involves hosting the model on-premises, in the cloud, or on edge devices, integrating it with software and workflows, and providing access to users or other applications via APIs.
Deployment also includes testing, monitoring, and managing updates to ensure reliable performance. The goal is to make the model fully operational and able to deliver consistent results.
Hosting
Hosting is a crucial part of deployment, as it’s the process of providing a place for an AI model to run so that it can be accessed and used. Models can be hosted on private servers, in the cloud, or on edge devices depending on the application. The choice of hosting affects how fast the model can respond, how many users or systems can access it, and how reliable it is, making effective hosting one in which the model is always available and ready to perform in tasks in real-world situations.
Access
Access is how users, software, or any other application interact with a trained AI model. In short, it defines the way a model is made available for use—such as through APIs, embedded services, or other interfaces.
For instance, a conversational tool might use an APIs to access a language model, while a business dashboard might embed a model to generate forecasts in real time. Access also involves managing permissions and controls so only authorized users or systems can use the model where needed.
On-premises deployment
On-premises deployment runs AI models on infrastructure fully managed directly by the organization, rather than in a third-party environment. This gives the organization full control over performance, data security, and compliance. Servers may be on-site or in a dedicated data center, while remaining under the organization’s responsibility and management, including maintenance, updates, and scaling.
Due to its nature, on-premises deployment offers high reliability and security, but requires significant resources and technical expertise.
Edge deployment
Edge deployment refers to the deployment of AI models directly on local devices such as sensors, cameras, wearables, or other Internet of Things (IoT) devices, instead of relying on remote cloud infrastructure.
This proximity reduces latency and allows for real-time analysis, making it suitable for scenarios like self-driving vehicles and smart home devices where speed and reliability are critical. Edge AI can work with limited or no internet connection. Due to these restrictive conditions, only specifically optimized models can be used for edge deployment.
Hosted models
Hosted models are AI models that run on infrastructure managed by a third party, such as a platform offered by a specialized vendor.
Responsibilities such as managing servers and technical environments are handled by the provider, and not the organization using the model. Users can access the model through APIs or web interfaces.
This setup reduces operational effort and simplifies scaling and maintenance, making hosted models chosen when ease of use and reliability matter more than full infrastructure control.
Cloud deployment
Cloud deployment is when a trained AI model runs on cloud infrastructure instead of on local or on-premises systems.
The model is hosted and operated using cloud service models such as Infrastructure as a Service (IaaS), Platform as a Service (PaaS), or Software as a Service (SaaS).
Cloud deployment is typically chosen for its flexibility, particularly in terms of scalability, cost efficiency, and ease of operation.
IaaS (Infrastructure as a Service)
IaaS is a way of providing virtualized computing resources, including virtual servers, storage, and networking over the cloud. With IaaS, providers deliver the underlying infrastructure required for AI deployment, (including maintenance, patches, upgrades, and troubleshooting), allowing customers to concentrate on software and strategic initiatives rather than managing physical hardware.
PaaS (Platform as a Service)
PaaS builds on top of IaaS and adds a layer of development and management tools and environments on top of IaaS providing a ready-to-use environment. It allows teams to build, train, and deploy AI models more efficiently without worrying about the underlying hardware. PaaS is ideal for organizations that want speed and simplified operations while still retaining control over the AI workflow.
SaaS (Software as a Service)
SaaS provides a ready-made software application over the internet. Here the provider hosts and manages everything: servers, updates, security, and maintenance.
Users don’t have to install anything or manage infrastructure—they just access the service via a browser or API.
This approach makes it fast to roll out, easy to scale and low-maintenancelow maintenance for customer organizations.
Models as a service (MaaS)
Models as a Service (MaaS) offers access to pre-hosted AI models through APIs or web platforms, without requiring organizations to manage infrastructure or deployment.
Users can integrate these models into their applications or workflows and immediately use their capabilities, while the provider handles hosting, scaling, updates, and maintenance.
Essentially, MaaS makes AI models ready-to-use services that can be deployed quickly and at scale.
