What Is a Foundation Model?
Last updated: 2026-05-18
What Is a Foundation Model?
A foundation model is a large AI system trained on massive amounts of data to perform a wide range of tasks. It's the underlying model that powers most AI tools you interact with. When you use ChatGPT, Claude, or Gemini, you're working with a foundation model (or a product built on top of one).
The model underneath a tool largely determines what it can and can't do. Knowing which model is powering something helps you make smarter choices about what belongs in your stack.
Definition and Examples
The term was popularized by Stanford's Institute for Human-Centered AI in 2021. Foundation models are:
- Pre-trained on enormous datasets (text, code, images, or combinations of all three)
- General-purpose — capable of many tasks without task-specific training
- Adaptable — can be fine-tuned or prompted for specific use cases
Major foundation models as of 2026:
- OpenAI: GPT-4o, GPT-4.5, o1, o3 — text and multimodal
- Anthropic: Claude Opus 4.7, Claude Sonnet 4.6, Claude Haiku 4.5
- Google: Gemini 2.5 Pro, Gemini 2.5 Flash — multimodal (text, image, audio, video)
- Meta: Llama 4 — open-weight, widely used for self-hosting and fine-tuning
- Mistral: Mistral Large, Mistral Small — strong on efficiency and cost
- Others: DeepSeek, Qwen, Cohere — competitive in specific niches
How Foundation Models Are Trained
Foundation models are built through pre-training: the model learns patterns from huge datasets, often billions of tokens of text or millions of images. This is expensive and only done by large labs. After pre-training, models are often refined with:
- Reinforcement Learning from Human Feedback (RLHF) — humans rate outputs to steer behavior
- Constitutional AI — principles baked in to reduce harmful outputs
- Instruction tuning — training on question-answer pairs so the model follows instructions
You don't train foundation models yourself. You use them via an API or a product built on one, and adapt them with prompts, fine-tuning, or RAG.
Concepts That Affect Tool Selection
Parameters — The learned weights inside the model. More parameters generally mean more capacity, but also higher cost and latency. A "70B" model has roughly 70 billion parameters. Model size affects quality and price.
Context window — How much input the model can consider at once (measured in tokens; roughly 4 characters per token in English). A 200K context window holds a long document; a 1M window can hold entire codebases. Tools built on models with larger context windows can handle longer conversations and bigger files.
Training data cutoff — Models have a knowledge cutoff date. A model trained on data through mid-2025 won't know about events from 2026. For time-sensitive tasks, check the cutoff and consider tools that use RAG to pull in fresh information.
Why This Matters for Tool Selection
When you browse the Model Directory, you'll find tools built on different foundation models. A writing assistant powered by Claude may excel at nuanced prose; one powered by GPT might be stronger at structured output. A coding tool might use a model optimized specifically for code. The underlying model affects:
- Capability — what tasks the tool can handle well
- Cost — token-based pricing ties directly to model choice
- Speed — smaller or optimized models respond faster
- Compliance — some models and providers offer better data privacy guarantees
Use the directory's filters and tool profiles to see which foundation model powers each tool. That helps you compare options and avoid paying for more (or less) than you need.