Runtimes

Most of these wrap the same engine, so the interesting differences are not speed but capability — which formats load, which memory levers you get, and whether it is built for one person or for many.

Runtime	Formats	KV quant	CPU offload	MoE offload	Built for
llama.cpp	GGUF	yes	yes	yes	Single-user focused
Ollama	GGUF	yes	yes	—	Fine for a handful of concurrent requests, not for a production workload
LM Studio	GGUF, MLX	yes	yes	—	Local server for personal use
vLLM	safetensors, AWQ, GPTQ, FP8, compressed-tensors	yes	—	—	This is the entire point
MLX	MLX (safetensors-based)	yes	—	—	Single-user focused

Head to head

lm studio vs ollama vllm vs ollama llama cpp vs ollama llama cpp vs vllm mlx vs llama cpp lm studio vs llama cpp