The local model runner feature matrix

Under the hood these are closer than they look. llama.cpp is the inference engine, and the others are, to varying degrees, wrappers around it or around the same techniques. So raw speed and model support are broadly similar on the same hardware. What actually differs is the experience: how you find and load a model, how much you configure, and how much is done for you.

If you've never done this before - LM Studio. It's a real desktop app with a searchable model browser that flags which quantizations your machine can handle. You click instead of type. Nothing else here is as forgiving on day one.

If you live in a terminal - Ollama. One command pulls a model, and the server is already running in the background. The workflow is tight and scriptable, and the defaults are sensible. The one gotcha for a browser app like mi·do·na is setting OLLAMA_ORIGINS so requests are allowed.

If you want the knobs - llama.cpp. You manage the GGUF file and the server flags yourself, which is exactly the point. It's the reference implementation, it moves fast, and permissive CORS is on by default so the browser reaches it without extra setup.

If you're here for roleplay - koboldcpp. A single-file executable with a simple launcher and a long history in the interactive-fiction community. It bundles its own web UI, serves an OpenAI-compatible endpoint, and turns CORS on by default.

Start with LM Studio if you want the gentlest on-ramp. Move to Ollama once you're comfortable at the command line. Reach for llama.cpp when you want full control, and koboldcpp if the roleplay tooling and presets appeal. There is no wrong pick: whichever you choose, mi·do·na talks to it the same way.

Ports and defaults reflect each project's current documentation as of July 2026. Check the official docs if a release has moved since.