Category

Where the models actually run

Training gets the headlines, but inference is where AI meets users: every answer is a model running on real hardware somewhere. These entries map the options - hosted provider APIs, self-hosted serving engines like vLLM, and local runners like Ollama - along with the techniques that make serving fast and affordable: batching, prompt caching, streaming, and structured output.

Entries in this category

Where this layer fits

Inference is the delivery layer of the stack. Upstream, models define what is being served and core concepts like tokens and quantization explain the raw mechanics. Downstream, evals and observability tell you whether all that serving machinery is producing good answers at a sane cost.

Not sure where to start? Inference providers frames the hosted-versus-self-hosted decision that everything else here builds on. Or browse everything at once in the A-Z index.