AdAI
Definition

Foundation Model is a large AI model trained on broad data at scale that can be adapted to many different tasks. ChatGPT, Claude, Gemini, Llama, and Mistral are all foundation models. So are image models like DALL-E and Stable Diffusion, audio models like Whisper, and multimodal models that handle text, images, and audio together. The defining feature is that one model can be applied to many tasks, rather than each task needing its own purpose-built model.

Key Takeaways

  • Foundation models are large AI models trained on broad data that can be adapted to many downstream tasks. The term was coined by Stanford CRFM in August 2021.
  • All current frontier language models (GPT, Claude, Gemini, Llama, Mistral, DeepSeek) are foundation models. So are major image, video, and audio generation models.
  • The economics of AI applications changed once foundation models became available. Building on top of one costs hours and tens of dollars. Training a comparable model from scratch costs tens to hundreds of millions.
  • SMBs almost never build their own foundation model. The usual path is to use one (via API or built-in features) or to fine-tune one on your specific data for a few hundred to a few thousand dollars.
  • Choosing a foundation model is a matter of matching capability, cost, and data sensitivity to the task. Frontier models for hard work. Smaller models for high-volume routine work. Self-hosted open models for sensitive data.
2021
the year Stanford CRFM coined the term "foundation model" in the report "On the Opportunities and Risks of Foundation Models", co-authored by more than 100 researchers
Source: Bommasani et al., Stanford Center for Research on Foundation Models, August 2021
16,000
NVIDIA H100 GPUs used for approximately 54 days to pretrain Meta's Llama 3 405B foundation model, one of the largest training runs publicly disclosed
Source: Meta AI, Llama 3 Technical Report, 2024

In Simple Terms

Before foundation models, building an AI system for a specific task usually meant training a specific model from scratch. Want to translate English to French? Train a translation model. Want to detect spam? Train a spam classifier. Want to summarise documents? Train a summariser. Each model knew one thing, did one job.

Foundation models broke that pattern. One large model, trained on enormous amounts of broad data, can do all three of those jobs and dozens of others, including ones the makers never specifically planned for. GPT-4 was not designed to write Python code. It does so anyway, well, because the training data included a lot of code. The same model writes contracts, summarises meetings, classifies tickets, answers customer questions, and translates languages. The shift is from "one model per task" to "one model that adapts to many tasks."

For an SMB, this matters because every AI feature you now see embedded in your business tools (HubSpot AI, Notion AI, Microsoft Copilot, Intercom Fin, Otter, Fathom) is built on top of a foundation model. The tool's makers did not train the AI themselves. They wrapped a foundation model with their own data, prompts, and interface.

The Major Foundation Models in 2026

The frontier of the field, broken into rough tiers.

Frontier closed models. OpenAI's GPT-5 and GPT-4o family. Anthropic's Claude Opus 4 and Sonnet 4. Google DeepMind's Gemini 3. These are the most capable models available, accessed via paid API or consumer subscription. Not open: the weights and training data are confidential.

Open-weight models. Meta's Llama 3 and Llama 4. Mistral's Mistral Large and Mixtral series. DeepSeek's V2 and V3. These are released with weights public, so a business can download and run the model on its own hardware. Useful when data sensitivity rules out sending requests to a third-party API. Llama 3 405B's published training compute (16,000 NVIDIA H100s for about 54 days) gives a sense of scale.

Smaller efficient models. GPT-4o-mini, Claude Haiku, Gemini Flash, Llama 3 8B, Mistral Small. Trade some capability for far lower cost and faster inference. Often sufficient for routine business work (classification, simple drafting, structured extraction) at a fraction of the price of frontier models.

Specialised foundation models. Image (DALL-E 3, Midjourney, Stable Diffusion, Flux). Video (Sora, Veo, Runway, Luma). Audio (Whisper, ElevenLabs models). Each is a foundation model within its modality. Many SMB tools wrap these for specific tasks (image generation in Canva, video in Descript, transcription in Otter).

How an SMB Uses Foundation Models in Practice

Three patterns, in order of how common they are.

Through the tools you already use. HubSpot AI runs on a foundation model. So do Microsoft Copilot, Notion AI, Intercom Fin, ChatGPT Team, Claude for Work. You pay for the tool. The foundation model is the engine the tool runs on. No technical decisions required.

Through a direct API or chat product. ChatGPT, Claude.ai, Gemini, or the corresponding APIs. An SMB owner uses ChatGPT directly for drafting emails or brainstorming. A custom AI agent built in n8n, Make, or a custom app calls OpenAI's or Anthropic's API directly. This is the configurable middle ground.

Through fine-tuning or self-hosting. Less common for SMBs, but increasingly accessible. OpenAI, Anthropic, and Google all offer fine-tuning APIs that let you train a smaller foundation model further on your own data for a few hundred to a few thousand dollars. Self-hosting an open-weight model on your own server or private cloud is the privacy-maximising option, usually relevant for healthcare, legal, or financial SMBs with strict data residency rules.

“We call these models foundation models to underscore their critically central yet incomplete character.”

Rishi Bommasani, Lead author, On the Opportunities and Risks of Foundation Models, Stanford CRFMCenter for Research on Foundation Models report, Stanford University, August 2021

Frequently Asked Questions

Is a foundation model the same as a large language model (LLM)?
Overlapping but not identical. All current LLMs (GPT, Claude, Gemini, Llama, Mistral) are foundation models. But the foundation-model category also includes large vision models (Stable Diffusion, DALL-E, Midjourney's underlying model), audio models (Whisper, ElevenLabs models), and multimodal models that combine all three. LLM is the language subset. Foundation model is the broader family.
Why are foundation models such a big deal?
Two reasons. One: capability. A single foundation model can do dozens of tasks (write, summarise, translate, classify, generate code, answer questions) that previously needed dozens of separate specialised models. Two: leverage. Building on top of an existing foundation model takes hours and tens of dollars. Training a comparable model from scratch costs tens to hundreds of millions of dollars. The economics of building AI applications completely changed once foundation models became available.
Do I need to build or train my own foundation model?
No. Essentially no SMB does, and almost no enterprise does either. Foundation models are built by frontier labs (OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek) at enormous cost. SMBs use them as customers, through APIs or built-in features. The few cases where a business builds its own foundation model are limited to large enterprises in regulated industries with very specific needs (Bloomberg's BloombergGPT, JPMorgan's IndexGPT, defence-sector models).
What is the difference between using a foundation model and fine-tuning one?
Using means calling the model as-is, possibly with a prompt that tells it what role to take. Fine-tuning means taking an existing foundation model and continuing to train it on your own data so it gets better at your specific use case. Fine-tuning typically costs a few hundred to a few thousand dollars (vs. tens of millions for original training) and is now a few clicks away on OpenAI, Anthropic, and Google's platforms.
Which foundation model should an SMB use?
Depends on the task. For general text work (writing, summarising, analysis), the major frontier models (GPT-5, Claude 4, Gemini 3) are all capable; pick by price and how well it integrates with your existing tools. For high-volume tasks where cost matters, smaller models (GPT-4o-mini, Claude Haiku, Gemini Flash) usually do the job at a fraction of the cost. For tasks involving sensitive data you do not want to send to a third party, an open-weight model (Llama, Mistral) hosted yourself or by a private cloud is the cleaner answer.

Related Glossary Terms & Resources

Machine Learning

The broader field. Foundation models are a recent (2018 onwards) development within it.

Inference

How foundation models actually get used: by calling them for inference, usually through an API.

Embedding

A specific output type that foundation models produce. The numerical representation of text or images that powers semantic search.

AI Automation Statistics 2026

Adoption and market data across foundation models and the broader AI category.

Join 5,000+ SMB owners getting weekly AI agent insights

Subscribe Free