A survey of self-hosted LLM inference orchestrators as of September 2026, comparing LocalAI, exo, GPUStack, and vLLM across features like multi-machine support, modalities, and cache-aware routing. LocalAI offers breadth with distributed mode and multiple modalities; exo optimizes Mac clusters with tensor parallelism; vLLM and llama.cpp serve as underlying engines for other systems. The choice depends on hardware, scale, and whether single-model or multi-model inference is needed.