alapmodell (FM)
in English: foundation model
A foundation model is a general-purpose AI model trained on vast amounts of data, serving as a base for various applications from chatbots to image generators.
In plain terms
Imagine a freshly graduated person with very broad knowledge: they have read extensively, understand a little about many things, but have never worked at a single job. The base model is like this: it has general knowledge, and specific tasks can be built upon it.
Previously, a separate AI was typically trained for each task: one for translation, one for spam filtering, one for image recognition. The base model is trained once, very expensively, but afterward the same model can write letters, summarize text, code, or analyze images.
This is why it is called a foundation: like the foundation of a house upon which others build. Most AI products you encounter do not use their own model but are built on a few major companies' foundation models or on freely downloadable open models.
How it works
- Pretraining: the model processes vast amounts of text, images, code, or audio while practicing a simple task, such as predicting the next word.
- Learning patterns: during extensive practice, it does not receive rules but automatically filters out the patterns of language, facts, and relationships, which are fixed in the model's internal numbers (parameters).
- Customization: a ready-made model can be adapted to a specific task through fine-tuning or simply well-written instructions, requiring only a fraction of the data and money.
- Integration: developers embed the model into applications, such as customer service chatbots, search engines, or agents that independently execute multi-step tasks.
- Usage: when you ask a question, the model generates the answer based on learned patterns; its knowledge typically extends only to the training cutoff date unless it receives fresh sources.
An everyday example
You ask the chat app on your phone to draft a polite complaint letter to your service provider, then, in the same app, you photograph your electricity bill and ask what a line item on it means. Both requests are handled by the same foundation model. Chances are that a similar model also runs behind your bank's customer-service chatbot.
Who is affected
A common misconception
Limits and risks
- It can state falsehoods in a convincing tone (this is called hallucination), so always verify important facts from other sources.
- Its knowledge is limited to the time of training, so it can only reliably report on recent events if it receives a search tool or updated document.
- It can adopt biases and prejudices from training data, which may then appear in every application built on top of it.
- Training is so expensive that only a few players can afford it, so many products and companies depend on the same few providers.
Similar, but not the same
- large language model – The large language model is the text-working type of foundation model; however, foundation models can also be created for images, audio, video, or even protein structures.
- chatbot – A chatbot is a finished product with a user interface and rules, while the base model is the engine running inside it.
- general AI – The foundation model is versatile but does not match human thinking in every field; artificial general intelligence is currently a goal, not an existing product.
- base model – A base model is the pre-trained, not yet fine-tuned version of a foundation model that only continues text; models used in chatbots are further trained from this to follow instructions.
Related concepts
large language model, multimodal model, image generation model, open-weight model
generative AI, fine-tuning, pre-training, transformer, AI-agent
A brief history
The term 'foundation model' was introduced by researchers at Stanford University in 2021 in a widely discussed study. The EU's AI Act, adopted in 2024, uses the term 'general-purpose AI model' in a similar sense.
Where things stand
By the end of 2026, the vast majority of everyday AI tools will be built on a few large foundation models, and leading models will handle images and audio alongside text, with an increasing number also processing video. In the EU, providers of general-purpose AI models have been subject to specific obligations since August 2025, compliance with which the European Commission can enforce with fines from August 2026.
In the news
- India releases its own foundation models as South Korea sets a record budget for sovereign AINov 2, 2026 · Agent Age
- ChatGPT could soon analyse Health Connect and Google Health data on AndroidNov 2, 2026 · Health
- Nvidia unveils Cosmos foundation models and Omniverse digital twin upgrades for industrial automationOct 21, 2026 · Ipar
- AI Act obligations for general-purpose AI models: biological risk and incident reportingOct 19, 2026 · Jog
- GPAI Code of Practice transparency chapter published as Commission readies guidelinesSep 19, 2026 · Jog
- FDA clears RadNet's foundation model-based chest X-ray AI toolSep 17, 2026 · Health
- Skild AI reaches $100 million in annual recurring revenue after 10 monthsSep 14, 2026 · Ipar
- Skild AI uses NVIDIA physical AI to teach robots new tasks from a single videoSep 10, 2026 · Ipar
Also known as
alapmodell · foundation model · Foundation model · general-purpose AI model · general-purpose AI model · GPAI-modell · FM
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.