A giant file of numbers. Billions of learned parameters, or "weights" (GPT-2 is 6.4 GB).
Training writes the weights. Done once. Costs up to hundreds of millions in compute.
Inference runs the weights to answer your question, on every query. That's the ongoing cost of AI.
Why GPUs? A model is mostly enormous matrix math. GPUs do thousands of multiplications in parallel and need very fast memory to keep the weights flowing.
Flavors of models
Text / LLMsImage generationSpeech & audioVideoMultimodal (all of the above in one)
Open vs closed weights
Open: the weights are published, so anyone can download and host them, often via Hugging Face (Llama, DeepSeek, Qwen, Mistral).
Closed: only the maker serves it, via API (GPT, Claude, Gemini).