Each stack page lists providers with a ready-made deployment path, GPU sizing by model class, and today's rental floors.
vLLM is the most widely deployed open-source inference engine, built around PagedAttention for high-throughput serving. It exposes an OpenAI-compatible API, supports tensor parallelism across multiple GPUs, continuous batching, and quantized checkpoints (FP8, GPTQ, AWQ). If you are serving an open model to real traffic and don't have a reason to pick something else, you pick vLLM.
SGLang matches or beats vLLM on raw throughput for many workloads and pulls ahead when requests share prefixes (agents, RAG, multi-turn chat) thanks to RadixAttention prefix caching. It also has first-class structured-output (JSON schema) support. DeepSeek and several inference providers run SGLang in production.
Ollama wraps llama.cpp with a one-line install, a model registry, and an OpenAI-compatible API. It runs GGUF quantized models efficiently on a single GPU (or CPU), making it ideal for development, internal tools, and low-traffic apps. It is not built for high-concurrency production serving. That's vLLM/SGLang territory.
TGI is Hugging Face's inference server, powering their own Inference Endpoints. It offers continuous batching, quantization support, and guided decoding. Choose it when you deploy through Hugging Face infrastructure or want their supported path; elsewhere the ecosystem has largely consolidated on vLLM.
Not sure which? Start from the breakeven calculator or the GPU price tables.