AI concepts
Foundation model (a base AI model)
A large machine learning model trained on a vast, diverse dataset, designed as a universal starting point for adaptation to specific tasks. The term was introduced by Stanford CRFM in 2021 as a counter to narrow specialist models. Current foundation models include GPT, Claude, Gemini, Llama, and Mistral.
Primary source: Stanford CRFM, Anthropic, OpenAI Technical Report 2024
The concept was introduced in the manifesto "On the Opportunities and Risks of Foundation Models" by the Stanford Center for Research on Foundation Models in August 2021. The authors (including Rishi Bommasani, Percy Liang, and Dan Jurafsky) argued that training a model once and using it many times for different tasks changes the economics of AI in a way comparable to the introduction of the LAMP stack in web hosting.
Four defining traits
The Stanford CRFM Foundation Models Report lists: the scale of parameters (billions to trillions), training on multimodal datasets (text, code, image, audio), generality of application (one model for many tasks), and adaptability through fine-tuning or prompting.
Three generations
Anthropic Technical Report 2024 describes the evolution. The first generation (2018 to 2022): BERT, GPT-3, T5, narrow capabilities. The second generation (2022 to 2024): GPT-4, Claude 2, Gemini 1, multimodal capabilities. The third generation (2024 onward): GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Pro, reasoning capabilities, tool use, agentic behaviours.
The economics of foundation models
OpenAI Technical Report Q4 2024 reports the cost of training a GPT-4 class model in the range of 80 to 100 million dollars. Llama 3.1 405B (Meta, released as open-weights) is estimated at 60 to 90 million. This capital scale means that foundation models are built by 5 to 10 firms globally, and the rest of the market is a consumer.
For a Polish enterprise the practical consequence is: a foundation model is bought through an API, not built. The strategic decision is not "which model to train", it is "which providers to choose and how to avoid vendor lock-in".
Open-weights vs proprietary
In 2025 the ecosystem split into two camps. Proprietary (OpenAI GPT, Anthropic Claude, Google Gemini): higher quality, higher price, no possibility of local hosting. Open-weights (Meta Llama, Mistral, DeepSeek, Qwen): lower benchmark quality, much lower operational cost, the option of on-premise deployment.
For regulated firms (financial, medical, legal), open-weights are becoming the standard, despite lower benchmark quality, because the data never leaves the client's infrastructure.
We make the decision on which foundation model to use for your company's specific use case within the AI Readiness Audit.