presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Foundations /Foundation model

    Foundation model

    Also called: base model.

    Detail

    In five seconds

    What it is

    A machine-learning model pretrained at large scale on broad data can be adapted for many tasks without training a new model from scratch.

    In plain words

    It is a large starter model trained on broad data so people can adapt it for many different jobs later.

    An everyday example

    A musician who has spent years learning rhythm, harmony and many styles does not begin again from the first lesson whenever a new performance is booked. That broad training helps them learn a jazz set, accompany a choir or play at a school concert. Each performance still requires the right music, rehearsal and adjustment to the audience. In the same way, a foundation model supplies broad capabilities that can be adapted for different tasks, but it is not ready for every task without further work and checking.

    Where the story stops working

    A library catalogue is a human-made index with explicit entries. A foundation model does not store knowledge in neat labelled cards. It learns statistical patterns through numerical adjustment, and it can generate fluent but wrong output. Also, a catalogue mostly helps you find existing books, while a foundation model can also produce new text, images or other outputs.

    Remember thisA foundation model is a reusable starting point for many AI applications, not a finished product. Every application still needs adaptation, testing, safeguards and oversight.

    This term is contested

    The term has no single universally accepted boundary. Different authors use it more narrowly or more broadly, especially when distinguishing it from large language models, multimodal models and frontier models.

    Understand it

    Understand it

    Why it matters

    Non-specialists hear this term in procurement, governance and risk discussions because it changes how AI systems are built and bought. Instead of commissioning a separate model for each use, organisations may start from one broadly trained model and adapt it. That can reduce time and cost, but it also changes the questions leaders should ask: what data was used in pretraining, what limits are known, what tasks has it actually been evaluated for, who controls the weights or access, and what safeguards exist for privacy, bias, safety and accountability. The label also matters because not every impressive model is a suitable foundation for public-sector work.

    How a foundation model becomes useful for many tasks

    After viewing, the reader should be able to explain that one broadly pretrained model can be adapted into several task-specific systems.

    1. 01

      Large varied data

      Training starts with a very large dataset, often containing text, images, audio or several types together.

      Draw a leftmost rounded rectangle labelled 'Large datasets' with small icons for document, image and audio wave; add a short note 'broad, varied data'.

    2. 02

      Pretraining

      The system adjusts many parameters to learn broad statistical patterns in the data.

      Draw an arrow to a second box labelled 'Pretraining'; inside, show gears or layered nodes with a caption 'learns patterns'.

    3. 03

      Base model produced

      The output is a reusable starting model, not yet tailored to one organisation's exact task.

      Draw an arrow to a central highlighted box labelled 'Foundation model / base model'.

    4. 04

      Adapt for tasks

      The same starting model can then be fine-tuned, prompted, evaluated or connected to tools for different uses.

      From the central box, draw three arrows to smaller boxes labelled 'summarise documents', 'classify images' and 'search assistant'; add small checkmark icons near each to suggest separate evaluation.

    Show one left-to-right flow with four boxes. Large mixed data goes into pretraining, which produces a base model. From that single model, three branches lead to different adapted uses such as summarising, image labelling and search assistance. Read from left to right, then down the branches.

    On hover or tap, each box reveals a one-sentence note about inputs, process or outputs; selecting a branch highlights that one base model can support several downstream uses.

    Describe this without the diagram

    The diagram is a left-to-right flow. First, a box labelled 'Large datasets' shows that training begins with broad data such as text, images and audio. An arrow leads to 'Pretraining', where the model learns statistical patterns by adjusting parameters. Another arrow leads to a central box labelled 'Foundation model or base model', showing the reusable result of that first training stage. From there, three arrows branch outward to 'summarise documents', 'classify images' and 'search assistant'. The main message is that one broadly pretrained model can be adapted for several different tasks instead of training a separate model from zero for each task.

    With reduced motion: A static four-stage flowchart with one central branch point. No animation. The branch lines are always visible to show one pretrained base leading to several later uses.

    A foundation model is a machine-learning model trained first on a very large and varied collection of data. That first stage is often called pretraining. The aim is to build a general starting point that can later be adapted to many tasks.

    What goes in at the start is usually large-scale data, such as text, images, audio or a mix of these. During training, the system adjusts many numerical settings, often called parameters, so that it becomes better at predicting patterns in that data. What comes out is not a finished service by itself, but a reusable base that others can fine-tune, instruct, connect to tools or place inside an application.

    This is different from a model built for just one narrow job, such as detecting a specific defect in one factory or forecasting one type of demand. A foundation model is intended to be general-purpose within the kinds of data it was trained on. It may support summarising, classification, drafting, search assistance or other tasks after adaptation, but that does not mean it is equally good or safe for all of them.

    People often confuse foundation models with large language models. Many large language models are foundation models, but not all foundation models work only with language. Some handle images, speech or several data types together. The exact boundary is still debated, so the term is useful but not perfectly settled.

    How “A musician learning a new set list” maps onto the real thing

    In the storyIn the AI system
    the large catalogue built from many bookspretraining on a very large and varied dataset
    different readers asking different questionsmany downstream tasks such as summarising, classifying or searching
    the same catalogue reused instead of rebuilt each timea base model adapted rather than training a new model from scratch
    library staff checking resultshuman oversight, evaluation and task-specific safeguards

    A worked example

    A reporting officer at a regional bureau is asked to pilot an assistant that produces first drafts of a weekly monitoring report from approved public material in English, Arabic and French.

    Starting from: A broadly pretrained multilingual model, an approved collection of public reports and human-reviewed examples showing the required structure and terminology.

    1. The officer starts with the pretrained model rather than commissioning a model trained from scratch for weekly reporting.
    2. The pilot connects the model to the approved document collection and gives it instructions based on the established monitoring report format.
    3. Reviewers test the assistant on material withheld from development, checking factual accuracy, omissions, citations and performance in each language.
    4. The assistant is limited to first drafts. The reporting officer verifies every claim against the source material before the report enters clearance.

    Result: One broadly pretrained model has been adapted to a specific reporting task without being rebuilt from zero. The same model could support other applications, but each would require separate configuration, evaluation, controls and human review.

    In a United Nations workplace

    At a country office town hall, a staff member asks whether the proposed drafting assistant is a different AI model from the one demonstrated for document search. The knowledge management officer has to answer before the discussion moves to the next agenda item.

    Both applications may use the same foundation model. Its broad pretraining provides a reusable starting point, while prompting, retrieval or further tuning adapts it for drafting or search. Each application still needs its own design, testing and oversight.

    Before you try thisReuse does not make every use appropriate. The officer should not enter personal, protection-sensitive, security-sensitive or other confidential material into a tool that has not been approved for it. Each application also requires checks for accuracy, access controls, language performance and fitness for purpose.

    Where you may hear it

    • A vendor demo during a procurement meeting
    • An ICT governance board reviewing an AI pilot
    • An inter-agency working group discussing shared digital tools
    • A front-office briefing before clearance of an AI concept note

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • A foundation model can do any task well straight away.

      Broad pretraining gives a useful starting point, not guaranteed competence. Performance varies by task, language, domain and context. A model may need further adaptation, testing and safeguards before it is reliable enough for real use.

    • Foundation model just means chatbot.

      Many chatbots are built on foundation models, but the terms are not the same. A chatbot is an application or interface. A foundation model is the underlying broadly pretrained model, which may support chat, search, image tasks or other uses.

    • If a model is called foundational, its output is authoritative.

      The term describes how the model was trained and reused, not whether each answer is correct. Outputs still need verification because models can reflect gaps in data, produce errors and perform unevenly across topics or languages.

    Quick check

    Which description best fits a foundation model?

    Go deeper

    Go deeper

    The simple version

    It is a large starter model trained on broad data so people can adapt it for many different jobs later.

    The technical version

    A foundation model is a machine-learning model, usually with many parameters, pretrained on large-scale and diverse data to learn general-purpose representations or predictive capabilities that transfer across downstream tasks. After pretraining, it can be adapted through fine-tuning or instruction tuning, or applied through prompting, retrieval, tool use or other methods. The term usually implies broad reuse across applications, but definitions differ: some authors restrict it to models with substantial scale and wide task transfer, while others use it more loosely for any broadly pretrained base model. It is related to, but not identical with, large language models, multimodal models or frontier models.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    A foundation model is a reusable starting point for many AI applications, not a finished product. Every application still needs adaptation, testing, safeguards and oversight.