A foundation model is a large AI model trained on broad data at scale, designed to be adapted to a wide range of downstream tasks. The term distinguishes these general-purpose models from earlier AI systems that were trained specifically for a single narrow task (e.g., a model trained only to classify spam email).

A human silhouette facing a vast red-lit industrial tower: a single structure large enough to underpin many purposes.
A foundation model is the tower. It was built at enormous scale. You do not rebuild it. You stand in front of it, understand what it offers, and design your system around what it can do.

Foundation models are the architectural shift that made modern enterprise AI practical: instead of training a new model from scratch for each use case, you adapt a single pre-trained foundation model - through prompting, fine-tuning, or retrieval augmentation - to your specific application.

What Makes a Model a Foundation Model

Scale - Foundation models are trained on massive datasets (hundreds of billions to trillions of tokens) using large compute clusters. This scale is what produces general-purpose capability.

Transfer learning - The model learns general representations during pre-training that transfer to specific tasks. A model trained on general text can answer questions, write code, translate languages, and classify sentiment - tasks not explicitly trained for.

Adaptability - Foundation models are designed to be adapted. Through prompting (zero-shot, few-shot), fine-tuning, or grounding with external knowledge, they can be directed to specialized tasks without retraining from scratch.

Major Model Families

Anthropic Claude - Strong general reasoning and agentic coding, long context (1M tokens on current Opus and Sonnet models at the time of writing, September 2026), conservative safety profile. Available via the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.

OpenAI GPT series - Widely deployed, strong code and reasoning, extensive ecosystem of tools and integrations.

Meta Llama and Muse - Llama is the best-known open-weights family; weights can be downloaded and self-hosted, which is cost-effective for high-volume workloads where self-hosting infrastructure is feasible. Llama 4 (April 2025) is the most recent Llama release. Since 2026 Meta’s flagship is the closed-weight Muse Spark family, and its current open-weight model is Muse Glimmer.

Google Gemini - Multimodal from the ground up (text, image, audio, video). Available via Google Cloud Vertex AI (rebranded Gemini Enterprise Agent Platform in April 2026).

Mistral - European models with strong data residency story. Competitive performance on European language tasks.

Amazon Nova and Titan - Amazon’s own models, tightly integrated with AWS services. Amazon Nova (currently the Nova 2 generation) is Amazon’s generative model family; the older Titan Text Embeddings V2 model remains available on Bedrock and is widely used in RAG pipelines.

Specific model versions change every few months. For the current lineup, context windows, and prices, see the LLM landscape .

Multimodal Foundation Models

Newer foundation models extend beyond text to multiple modalities:

  • Text + image - Models that can process images and text together (describe an image, answer questions about a diagram, extract data from a chart)
  • Text + audio - Speech understanding and generation
  • Text + code - Models with specialized code understanding, often from combined text and code training

For enterprise AI, multimodal capabilities open use cases that pure text models cannot handle: document intelligence with embedded images, analysis of visual inspection data, voice interfaces.

Practical Implications

Foundation models changed the economics of enterprise AI. Before foundation models, building an AI system for a new use case often required collecting thousands of labeled examples and training a new model. Now, a capable AI system for a new task can often be built in days, using prompting and existing infrastructure.

The flip side is that foundation model behavior is harder to control precisely than a narrow task-specific model. Guardrails, output validation, and human review processes are necessary for production enterprise use.

Sources and Further Reading

  1. Bommasani, R. et al. (2021). “On the Opportunities and Risks of Foundation Models.” Stanford Center for Research on Foundation Models (CRFM). arXiv:2108.07258., The paper that coined the term “foundation model” and provided the first systematic analysis of capabilities, limitations, and societal implications. https://arxiv.org/abs/2108.07258
  2. Brown, T. et al. (2020). “Language Models are Few-Shot Learners.” NeurIPS 2020., GPT-3 paper demonstrating that large-scale pre-training produces in-context few-shot learning capabilities; the empirical foundation for treating LLMs as general-purpose AI systems. https://arxiv.org/abs/2005.14165
  3. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2018). “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.” arXiv:1810.04805., Established masked language modeling and fine-tuning as the dominant approach for adapting pre-trained language models to downstream tasks. https://arxiv.org/abs/1810.04805
  4. Touvron, H. et al. (2023). “Llama 2: Open Foundation and Fine-Tuned Chat Models.” arXiv:2307.09288., Meta’s open-weights foundation model family; the primary reference for open-access foundation models. https://arxiv.org/abs/2307.09288
  5. Anthropic. “Models overview” (accessed 25 September 2026). https://platform.claude.com/docs/en/about-claude/models/overview
  6. AWS. “Amazon Titan Text Embeddings models,” Amazon Bedrock User Guide (accessed 25 September 2026). https://docs.aws.amazon.com/bedrock/latest/userguide/titan-embedding-models.html