LocalAI CLI Cheatsheet - local-ai Self-hosted Inference Server Full Reference

Plenty of teams want to keep using the OpenAI SDK without shipping code off the network, and LocalAI is the swap-in path: point base_url at localhost and models, embeddings, and speech all stay on your own machine. This table groups service startup, model management, and API and fine-tuning, explaining the difference between run and serve under the local-ai entry point, how the models subcommands install and list what you have, and how the fine-tuning endpoints are called. The deciding variable is the match between hardware and model size: whether it runs without a GPU depends on parameter count and quantization level, not on LocalAI itself, so validate the whole chain with a few-hundred-megabyte quantized model before attempting anything larger. Reach for it when compliance requires data to stay inside the network, when a development environment must reproduce a model behavior offline, or when routing between a small local model and a large hosted one. Against GPU-only serving stacks such as vLLM, it behaves more like a local AI gateway with model management and multimodal support built in.