Artificial intelligence has become remarkably easy to use and surprisingly difficult to explain.
A person can open ChatGPT, Claude, Gemini or Kimi and begin working within seconds. Behind that simple interface, however, sits an expanding vocabulary of technical terms: large language models, tokens, parameters, reasoning, GPUs, context windows, open weights, fine-tuning, data centers and benchmarks.
The terminology is growing almost as quickly as the technology itself. Even people who use AI every day can struggle to maintain an overview.
This guide explains the most important concepts in practical language. It is written for people who understand what AI can do but want a clearer picture of how the technology works, how models differ and why companies are investing billions of dollars in the race to build better systems.
What is a large language model?
LLM stands for large language model.
At its core, an LLM is a mathematical system trained to recognize patterns in language. During training, it processes enormous amounts of text and repeatedly tries to predict what should come next.
Imagine the sentence:
The customer paid for the product at the…
Most people would expect words such as “checkout,” “register” or “counter” to follow. A language model performs a much more complex version of the same prediction, using patterns learned from a vast quantity of data.
When someone asks an LLM a question, the model does not normally retrieve one complete answer from a database. It generates its response step by step, predicting one small unit of text after another.
That apparently simple process can produce advanced capabilities, including writing, translation, summarization, coding, document analysis and certain forms of reasoning.
The model is therefore not a traditional database, and it is not a search engine. It is a system for generating outputs based on patterns it learned during training.
The model is not the product
One of the most important distinctions in AI is the difference between a model and an application.
GPT is a family of models developed by OpenAI. ChatGPT is a product that uses those models.
Claude is Anthropic’s model family, but the same name is also used for the company’s assistant.
Gemini refers both to Google’s models and to the application through which many people use them.
Kimi is an AI product from Moonshot AI, while Kimi K3 is the model that currently powers many of its capabilities.
A finished AI product may combine a language model with search, memory, databases, business applications, security controls, company documents and external tools. The LLM provides much of the language understanding and generation, but it is only one part of the complete system.
This distinction becomes particularly important when companies evaluate AI. A better user experience does not always mean that the underlying model is more intelligent. The application may simply have better tools, access to fresher information or a more effective interface.
The main types of language models
There is no single type of LLM. Models are developed and adapted for different purposes, levels of performance and operating environments.
Base models
A base model has completed the initial stage of training on large quantities of data. It has learned language structures, relationships, writing styles and many general patterns.
It may be highly capable, but it has not necessarily been optimized to behave like a helpful assistant.
A base model can be compared with someone who has read a vast library but has not yet learned how to conduct a structured professional conversation. It may continue text effectively without always understanding that the user expects an answer, a summary or a recommendation.
Base models are frequently used as foundations for further development.
Instruct and chat models
An instruct model has received additional training to follow human instructions.
It learns that a request to summarize a report should produce a summary, while a request to write an email should produce an email rather than simply continue the wording of the prompt.
Most public AI assistants use instruct or chat versions of a model. They are normally more useful in everyday work because they have been trained to respond in formats that people recognize and expect.
Reasoning models
Reasoning models are designed to devote more computational effort to difficult problems.
They may analyze several possibilities, divide a task into steps, use external tools and review parts of their own work before producing the final answer. This makes them particularly useful for mathematics, software development, scientific analysis, research and complex planning.
The trade-off is usually speed and cost. A model that spends more time and more tokens analyzing a problem may produce a stronger answer, but it will generally require more computing resources.
Multimodal models
A traditional language model works primarily with text. A multimodal model can process and sometimes generate several forms of information.
Depending on the model, this may include text, images, audio, video, documents, charts and source code.
A multimodal model might read a contract, analyze a photograph, interpret a diagram and answer spoken questions within the same interaction.
This is one reason the term LLM is gradually becoming incomplete. Some of the most advanced systems are no longer limited to language. They are becoming broader foundation models that understand and generate several types of content.
Google describes Gemini 3.5 as part of a model family combining advanced intelligence with action, while its Gemini Omni system is designed to work across multiple types of media. (Google DeepMind)
Specialist models
Some models are developed for a specific field or task.
They may focus on coding, mathematics, medicine, legal analysis, finance, translation, speech recognition or scientific research.
A specialist model can outperform a much larger general model within a narrow area because its training and evaluation are concentrated on that particular task.
This is an important lesson for businesses: the largest or most famous model is not automatically the best model for every project.
Small language models
Not every useful model needs hundreds of billions of parameters.
Smaller language models can be faster, cheaper and easier to operate. Some can run on company servers, laptops, smartphones or other local devices.
They may not match the strongest frontier models on complex research or reasoning tasks, but they can be highly effective for document classification, standard customer inquiries, local search, data extraction and repetitive operational work.
They can also offer important advantages in privacy, offline availability and control over where information is processed.
Agentic systems
An AI agent is not necessarily a completely new type of language model. It is usually a model connected to tools, memory, permissions and a defined process.
An agent may be able to search the web, read documents, use software, analyze spreadsheets, write code, send messages or complete a multi-stage business task.
The model provides language understanding, planning and decision support. The surrounding system gives it the ability to act.
This means that an agent’s quality depends on more than the model. It also depends on the tools it can access, the permissions it receives, the quality of its instructions and the controls that prevent unwanted actions.
Closed models, open weights and open source
The way a model is distributed can be as important as its intelligence.
Closed models
A closed model is normally accessed through a commercial application, API or cloud platform.
The provider operates the infrastructure and does not release the model’s internal weights. Customers can use the system, but they cannot normally download the complete model and operate it independently.
This approach can offer strong performance, easy access and reduced technical responsibility. The provider manages the infrastructure, model updates and much of the security.
The disadvantages can include dependency on the provider, limited transparency, changing prices and less control over deployment.
Open-weight models
An open-weight model makes its trained parameters available for download under a particular licence.
A company may be able to run it on its own infrastructure, adapt it, fine-tune it and continue using a selected version even when the original provider changes its commercial service.
This creates greater technical independence, but it also transfers more responsibility to the user. The organization must operate the infrastructure, secure the model, manage updates and evaluate its behaviour.
Open weights do not automatically make a model fully open source. The training data, complete training process or code required to reproduce the model may remain unavailable.
Open-source AI
The Open Source Initiative uses a broader definition. Under its Open Source AI Definition, users should have the freedom to use, study, modify and share the system, together with access to the materials needed to exercise those freedoms.
This is why many models commonly described as open source are more accurately called open-weight models.
The distinction is not merely academic. It affects transparency, independence, licensing, security reviews and the ability to reproduce or substantially modify the technology.
Self-hosted and local deployment
A self-hosted model runs in a company’s own data center, private cloud or controlled infrastructure.
This provides greater control over data, model versions, integrations and availability. It may be particularly valuable for organizations handling confidential information or operating under strict regulatory requirements.
Local deployment goes one step further by running the model directly on a device, such as a laptop, smartphone, vehicle or store system.
This can reduce latency and allow the AI to work without a permanent internet connection.
Hybrid deployment
Many companies are unlikely to choose one model or one deployment strategy for every task.
A powerful commercial model may be used for complex research, while an open-weight model handles sensitive internal documents. A smaller local model may support repetitive processes, while conventional deterministic software remains responsible for financial calculations or legally binding decisions.
The future will probably not be dominated by one universal model. It will be shaped by architectures that select the right model for each task.
What is a token?
A token is a small unit of information that a language model processes.
A token may be a complete short word, part of a longer word, a number, punctuation or another symbol. The exact division depends on the model and its tokenizer.
The word “store” may be represented as one token, while a longer or less common technical term may be divided into several.
Models receive input in tokens and generate output in tokens. This is why AI providers often calculate usage and pricing per million tokens.
Tokens are also used to measure the model’s context window.
What is a context window?
The context window is the amount of information a model can consider during one interaction.
It may contain the user’s current request, earlier messages, uploaded documents, tool results, system instructions and the answer being generated.
A larger context window allows a model to process longer documents and more extensive conversations. It does not, however, guarantee that the model will use every part of that information equally well.
The context window should also not be confused with permanent memory. Information inside the current context may disappear when the session ends unless the application stores and reintroduces it later.
Moonshot AI says its newly released Kimi K3 supports a context window of one million tokens, illustrating how model developers are competing not only on reasoning but also on the volume of information their systems can process at once. (Kimi)
What are parameters?
Parameters are the internal numerical values a model learns during training.
They influence how the model responds when it encounters particular patterns in its input.
A model described as having 70 billion parameters contains 70 billion learned numerical values. Larger parameter counts can provide greater capacity, but size alone does not determine quality.
Training data, architecture, post-training, efficiency and evaluation methods can matter just as much.
Some modern systems use a mixture-of-experts architecture. Such a model may contain an enormous total number of parameters while activating only part of the model for each token.
This is why parameter figures must be interpreted carefully. A model with trillions of total parameters is not necessarily using all of them for every response.
What is a GPU?
GPU stands for graphics processing unit.
GPUs were originally developed to handle the large number of calculations required to produce computer graphics. Their architecture allows them to perform many similar mathematical operations in parallel.
That same ability makes them highly effective for artificial intelligence.
A conventional central processing unit, or CPU, is designed to handle a wide variety of tasks. It is comparable to a small team of flexible specialists.
A GPU is more like a very large workforce performing thousands of similar calculations simultaneously.
Training a major language model requires an extraordinary number of mathematical operations. GPUs make it possible to divide that work across many processing cores and, at a larger scale, across thousands of connected machines.
What is an AI data center?
A data center is a physical facility containing servers, storage, networking equipment, electricity infrastructure and cooling systems.
An AI data center adds large numbers of GPUs and extremely fast connections between them.
The GPUs must exchange information continuously during training and model operation. This means that the networking architecture can be almost as important as the processors themselves.
AI infrastructure therefore involves much more than purchasing chips. It requires high-speed networks, storage, data pipelines, orchestration software, power, cooling, monitoring and security.
NVIDIA increasingly uses the term AI factory to describe infrastructure designed specifically for data ingestion, model training, fine-tuning and large-scale AI operation. The company describes tokens as one of the outputs produced by these facilities, with economics increasingly measured through factors such as tokens per second, tokens per watt and cost per token.
What does it mean to train an AI model?
Training means adjusting a model’s parameters so that it becomes better at producing the intended output.
The process normally involves several stages.
Pre-training
During pre-training, the model processes enormous amounts of data and repeatedly attempts to predict the next token.
When the prediction is inaccurate, the model’s parameters are adjusted slightly. This happens again and again across a very large number of examples.
The result is a base model that has learned language structures, general patterns and relationships from the material used during training.
Pre-training is normally the most expensive and infrastructure-intensive stage.
Post-training
Post-training teaches the model how it should behave when interacting with people.
The process may include examples of strong answers, instruction-following exercises, human preference comparisons, reinforcement learning, safety training and tool-use training.
Post-training helps transform a base model into a useful assistant.
It can also explain why two products built on related foundations may behave very differently. Their post-training, system instructions and safety policies may not be the same.
Fine-tuning
Fine-tuning adapts an existing model to a particular task, style or field.
A company might fine-tune a model to use its terminology, classify documents, follow a prescribed response format or understand a specialist workflow.
Fine-tuning normally begins with an already trained model. It does not mean rebuilding the entire system from the beginning.
RAG is not training
Retrieval-Augmented Generation, usually shortened to RAG, gives a model access to external information when it answers a question.
For example, the application may search a company knowledge base, retrieve relevant documents and place the most useful passages into the model’s context.
The model itself may remain unchanged.
The distinction is simple:
Fine-tuning changes the model. RAG changes the information available to the model.
RAG is particularly useful where information changes frequently. Laws, technical specifications, product information and internal policies can be updated in the knowledge source without retraining the complete model.
How is the quality of an LLM measured?
There is no single test that can identify the universally best language model.
Models are evaluated across a range of capabilities, including knowledge, mathematics, coding, reasoning, document understanding, tool use and multimodal performance.
Benchmarks
A benchmark is a standardized test containing questions or tasks that can be given to several models.
Benchmarks allow developers and researchers to compare performance under similar conditions.
However, benchmark results have limitations. A model may perform strongly on a test while being less reliable in a real business process. Test data can become familiar to model developers, evaluation methods can contain errors, and companies may select benchmarks that highlight their strongest capabilities.
OpenAI itself reported concerns in 2026 about the reliability and accuracy of SWE-Bench Pro, a widely used coding evaluation, illustrating why even popular benchmarks should not be treated as unquestionable measures of model quality.
Human evaluation
People can compare model answers and decide which is clearer, more useful, more accurate or better written.
This provides valuable information, but human preferences are not always objective. A confident and elegant answer may be preferred even when it contains a factual mistake.
Reliability
For business use, factual reliability may matter more than conversational style.
Organizations should test whether a model invents information, uses sources correctly, produces consistent answers and recognizes when it does not have enough evidence.
A model that performs well on general questions may still fail on a company’s specialist documents or operational terminology.
Speed and cost
The strongest answer is not always the most practical answer.
A model that requires several minutes and a large number of tokens may be appropriate for research but unsuitable for customer service, a retail checkout or a real-time operational process.
Companies must consider response time, token costs, infrastructure requirements, energy consumption and the number of simultaneous users the system can support.
The business test
The most valuable evaluation is usually the one built around the company’s own task.
A retailer evaluating AI for product classification should test real product records. A software company should use examples from its own repositories. A compliance organization should test the model against validated regulations and known interpretations.
The useful question is therefore not:
Which LLM is the best?
It is:
Which model performs best for this task, risk level, budget, response time and operating environment?
The companies competing to build the best models
The global AI race is no longer a contest between only two or three American technology companies.
OpenAI currently offers the GPT-5.6 family, including its flagship Sol model and lower-cost Terra and Luna variants. Google DeepMind is developing Gemini 3.5 and Gemini Omni, while Anthropic introduced Claude Sonnet 5 for coding, agents and professional work. SpaceXAI launched Grok 4.5 in July 2026, describing it as its strongest model for coding, agentic tasks and knowledge work.
A parallel race is unfolding around downloadable and more openly distributed models.
Meta has developed the Llama family, Google offers Gemma, Alibaba has built the Qwen ecosystem, DeepSeek released DeepSeek V4, and Mistral AI continues to develop both commercial and open-weight models. DeepSeek says V4 is available in Pro and Flash versions and has released model weights for community use.
Kimi K3 enters the race
One of the newest major releases is Kimi K3, introduced by Moonshot AI on July 16, 2026.
Moonshot describes Kimi K3 as a 2.8-trillion-parameter model with native visual capabilities and a context window of one million tokens. It powers the company’s chat product, AI agent and multi-agent system. Full model weights are scheduled for release on July 27, which would place Kimi K3 firmly within the competition around highly capable, openly distributed models.
The arrival of Kimi K3 is important for two reasons.
First, it shows that the frontier of AI development is becoming geographically and commercially broader. Strong models are emerging from a growing number of companies and ecosystems.
Second, it demonstrates that the competition is no longer based only on who can build the largest model. Companies are also competing on context size, multimodality, agent capabilities, coding, cost, speed and openness.
A model may lead in one category and lag in another.
The race is about more than intelligence
The companies building LLMs are competing across several dimensions at once.
They want models that reason more effectively, write better code, process longer documents, understand images, operate tools and complete tasks with less human involvement.
At the same time, they must make those models faster and less expensive to run.
Distribution has also become a strategic battleground. Closed-model providers compete through powerful cloud services and polished applications. Open-weight ecosystems compete by attracting developers, researchers and companies that want more control.
The result is not one AI market but several overlapping markets: models, cloud infrastructure, chips, applications, agents, developer tools and enterprise platforms.
A simple way to keep the overview
Modern AI can be understood as a stack.
Data provides the learning material.
Training adjusts the model.
Parameters contain the patterns the model learned.
GPUs perform the calculations.
Data centers provide the infrastructure.
Tokens are the units the model reads and generates.
The context window determines how much information it can consider at once.
Post-training teaches it how to respond and behave.
Tools and knowledge sources connect it to current information and real systems.
Applications turn the model into something people can actually use.
That combination, rather than the model alone, is where much of the practical value of AI will be created.
