Foundation Models are pre-trained large-scale AI models — primarily large language models (LLMs), but also multimodal (vision, audio, video), embedding, and reasoning models — accessible via hosted inference APIs, self-hosted weights, or managed platform SKUs. Core capabilities include text generation with extended context windows (100K+ tokens), multimodal input and output, function-calling and tool use, structured outputs, streaming responses, fine-tuning APIs, model-specific behavior tuning (system prompts, safety filters), and per-token pricing. The category is defined by a small set of frontier-model providers (OpenAI, Anthropic, Google Gemini, xAI, DeepSeek) and a growing set of open-weight model families (Meta Llama, Mistral, Alibaba Qwen) deployable on customer infrastructure.