[Official FAQ] Share your experience or follow-up questions below.

They solve slightly different problems, so the "best" one depends on what you are doing.

• Ollama: command-line first, very easy install, one command to pull and run a model, and it exposes a local API. Great for beginners and quick experiments.
• LM Studio: a desktop GUI where you can browse, download and chat with models, and optionally start a local server. Good if you prefer not to use the terminal.
• llama.cpp: the underlying inference engine for GGUF models, with the most knobs. It supports CPU, CUDA, Metal, Vulkan and more. Best when you want to tune performance or run on unusual hardware.
• vLLM: a serving framework focused on throughput and many concurrent users. It is mainly used on Linux with NVIDIA GPUs and typically loads Hugging Face weights rather than GGUF. Best for team or production serving.

Quick decision guide

• Just want to try models today: Ollama or LM Studio.
• Want maximum control or have unusual hardware: llama.cpp.
• Serving many users or building a product: vLLM.

Good news: most of them can expose an OpenAI-compatible endpoint, so your application code is usually easy to move between them. Differences in streaming, tool calling and structured output do exist, so test the features you rely on.

You can also run more than one. Many people prototype in Ollama or LM Studio and move to vLLM when they need concurrency.

What are you using now, and what made you choose it?