What are large language models (LLMs)?

What are large language models?

Large language models (LLMs) are AI systems that understand and generate human language – trained on billions of texts from the internet, from books and from specialist articles.

The German term is großes Sprachmodell. In practice, the English abbreviation LLM has become standard.

Distinction

How do LLMs differ from simple language models?

LLMs differ from simple language models in their size – billions instead of millions of parameters fundamentally change what a model can do.

Scale

Simple language models work with a few million parameters and recognise local patterns such as word pairs or short phrases. LLMs operate with hundreds of billions of parameters and thereby capture complex relationships across entire paragraphs.

Emergent abilities

Above a certain model size, abilities emerge that nobody explicitly trained. LLMs then solve logical tasks, write code or translate between languages – even though they only learned to predict the next word.

Context understanding

Simple models lose the thread after a few sentences. LLMs keep context windows of thousands to millions of tokens open and link information across long texts.

How it works

How do large language models work?

LLMs predict the next word – token by token, millions of times in a row. This simple principle produces texts that read as if a human had written them.

Tokenisation

The model breaks every text into small units: words, syllables or individual characters. These tokens form the computational basis for all further steps.

Transformer architecture

Since 2017, transformers – an architecture of neural networks – have processed all tokens in parallel rather than one after another. This architecture enables the training of models with hundreds of billions of parameters.

Self-attention

Self-attention weights the relationship between all words in a text – even across large distances. That is how the model recognises what a pronoun or a reference refers to.

Inference

For every request, the model calculates probabilities for all possible next tokens. It selects the most likely one and repeats the process until the answer is complete.

Training

How do developers train LLMs?

Developers train large language models in three phases – each stage refines language understanding and the quality of the answers.

Pre-training

The model processes billions of texts and learns linguistic patterns through self-supervised learning. It predicts masked or following words and adjusts its parameters over millions of passes.

Fine-tuning

After pre-training, developers specialise the model for certain tasks. A general language model learns, for example, to formulate more precise diagnoses from medical texts.

RLHF

Reinforcement learning from human feedback makes answers more helpful and safer. Human raters rate answers by quality, and the model learns from these ratings.

AI context

What distinguishes LLMs from conventional AI?

LLMs form a subfield of artificial intelligence – they focus on language, while AI as an umbrella term covers many different technologies.

Classic AI

Rule-based systems work with hard-coded decision trees and structured data. A spam filter checks emails against defined criteria – it does not learn on its own.

Machine learning

Machine learning algorithms recognise patterns in data without anyone programming the rules manually. An image recognition AI distinguishes cats from dogs by analysing thousands of example images.

LLMs as language specialists

LLMs process unstructured language, capture context and generate new texts – core tasks of natural language processing. They understand nuances, irony and ambiguous instructions – abilities that classic AI and simple machine learning do not cover.

Models

Which large language models are there?

Several providers develop large language models with different strengths – from multimodal processing and safety to open source.

ChatGPT

ChatGPT from OpenAI is one of the best-known LLM applications worldwide. The underlying GPT models process text, image and audio and achieve top scores in many benchmarks.

Claude and Gemini

Claude from Anthropic impresses with long context windows and a focus on safety. Gemini from Google integrates deeply into the Google ecosystem and combines search with language understanding.

Llama and open source

Llama from Meta follows an open-source approach and gives companies full control over data and infrastructure. There are also models such as Mistral from France or Granite from IBM.

Context

Is ChatGPT a large language model?

ChatGPT is based on a large language model but is more than just the model – OpenAI combines the LLM with a chat interface, safety filters and external tools.

The model

OpenAI's GPT series forms the core of ChatGPT. It generates text answers by calculating the most likely continuation token by token.

The interface

ChatGPT adds conversation history, user management and plugins to the model. These functions are not part of the LLM itself but of the product layer above it.

Safety layer

RLHF and content filters reduce harmful or wrong answers. Every output goes through several checks before the user sees it.

Application

What do companies use LLMs for?

Companies use large language models for tasks that previously required human language skills – from customer communication to internal knowledge work.

Customer support

AI chatbots for business answer customer enquiries around the clock in natural language. They automatically recognise complex cases and forward them to staff.

Content creation

LLMs draft blog articles, product descriptions and social media posts. Targeted prompt engineering controls the tone, length and format of the results.

Practice

How are LLMs changing work with text and data?

LLMs automate tasks that used to cost hours of manual work – from code creation to the analysis of entire document archives.

Code assistance

Developers use LLMs to write, review and translate code. Tools such as GitHub Copilot or Claude Code speed up the entire development process.

Translation

LLMs capture context and tone better than rule-based translation systems. They do not translate individual words but the meaning of entire paragraphs.

Summarising

An LLM condenses long reports, contracts or email threads to their key points in seconds. Teams thus save hours of reading time per week.

Data analysis

LLMs extract patterns and insights from unstructured text – for example from customer feedback, support tickets or market reports. Embedded in a workflow automation with AI, the model writes the results directly back into the CRM or BI dashboard.

RAG

What is retrieval-augmented generation?

Retrieval-augmented generation (RAG) connects an LLM to an external knowledge base – the model searches for relevant documents and bases its answer on their content.

Why RAG?

Without RAG, an LLM relies exclusively on its training knowledge. With RAG, it searches current company documents, manuals or databases in real time.

Three steps

The process follows a clear pattern: the system retrieves relevant text passages, augments the request to the LLM with them and generates a source-based answer.

Fewer hallucinations

RAG significantly reduces hallucinations because the model accesses verified information. Knowledge stays up to date – without expensive retraining of the entire model.

Grounding

What does grounding mean for LLMs?

Grounding anchors an LLM's answers in verifiable sources – the model speculates less and backs up its statements with concrete data.

Web search

The LLM searches the internet for current information and cites the sources. ChatGPT, Claude and Gemini already offer this function as standard.

Database connection

Companies connect LLMs to internal systems such as CRM, ERP or knowledge bases. The model then answers based on real business data instead of general training knowledge.

Tool use

LLMs call external tools – they calculate with calculators, query APIs or run code. That extends their abilities beyond pure text generation.

Fact-checking

Grounding mechanisms automatically compare statements with reference sources. The system flags uncertain information so that users can check it specifically.

Limits

What limits do large language models have?

LLMs calculate probabilities – they neither understand meaning nor check facts. That limits their reliability in critical applications.

Hallucinations

LLMs produce plausible-sounding statements that are factually wrong. Such hallucinations occur regularly, especially with facts, figures and source references.

Bias

Bias arises when training data over-represents certain perspectives. The model then reproduces social prejudices in its answers.

Energy consumption

Training a large language model consumes as much electricity as hundreds of households in a year. Ongoing operation also requires considerable computing resources.

Currency

Without grounding or RAG, an LLM only knows information from its training period. Current events, new laws or internal company data are missing.

Outlook

How are LLMs developing?

New architectures, more compact models and autonomous agents define the next phase of LLM development.

Multimodal models

LLMs no longer process only text but also images, audio and video in a single request. ChatGPT and Gemini already master this multimodal processing.

Small language models

Compact models such as Phi (Microsoft) or Gemma (Google) run directly on smartphones and laptops. They need no cloud connection and thus protect users' privacy.

Reasoning models

New architectures break complex problems down into traceable intermediate steps. They solve mathematical and logical tasks much more precisely than conventional LLMs.

Agents

What are AI agents?

AI agents are LLM-powered systems that solve multi-step tasks on their own – they plan, use tools and make intermediate decisions without human intervention.

Planning and execution

An AI agent breaks complex requests into sub-steps and works through them one after another. It decides for itself which tool to use for which step.

Tool use

Agents access APIs, databases, file systems and other software. They run code, search the internet or send emails – depending on the task.

Autonomy with control

Companies define guardrails within which the agent acts freely. That way they automate complex workflows without giving up control over critical decisions.

FAQ

Frequently asked questions about large language models