Ollama vs LM Studio vs llama.cpp: which local AI runner?

Search for a local AI runner and you get listicles ranking these three on speed, as though they were rival engines.

They are not rivals, and they are not three engines. Both Ollama and LM Studio run llama.cpp underneath. Ollama's repository pins a specific llama.cpp version in a file called LLAMA_CPP_VERSION and credits the project in its README. LM Studio's own pricing page states it runs "the latest local LLMs using llama.cpp and MLX".

So the tokens-per-second arguments are mostly noise. You are choosing a user interface and a licence, not an inference engine.

What each one actually is

	llama.cpp	Ollama	LM Studio
What it is	The inference engine	A CLI and local server wrapping it	A desktop GUI wrapping it
Interface	Command line, C/C++ library	Command line, HTTP API	Graphical app
Licence	MIT	MIT	Proprietary (free tier)
Language	C++	Go	Closed
Cost	Free	Free	$0 tier; paid cloud credits
Source available	Yes	Yes	No
Best at	Control, embedding, the newest features	Scripting, serving, automation	Browsing, trying, chatting

Star counts, as of July 2026: llama.cpp ~121k, Ollama ~176k. Ollama is more popular than the engine it runs on, which tells you most people want the wrapper, not the engine.

Why the licence is the real decision

This is the part the speed comparisons skip, and it is the one that can actually cause you a problem.

llama.cpp and Ollama are MIT licensed. You can read them, ship them, embed them in a product, and run them commercially without asking anyone.

LM Studio is proprietary. The app is free to download and there is a $0 tier, but it is a product governed by terms of service, not an open-source licence. That is not a criticism — it is a well-made app — but if you intend to build a commercial product on top of it, or deploy it across a company, that is a conversation with a vendor rather than a git clone. LM Studio sells Enterprise and Teams plans precisely for this.

If your use is personal, this does not matter at all. If it is commercial, it is the first thing to check and the last thing any listicle mentions.

Choosing

If you…	Use
Want to try local models without reading a manual	LM Studio
Want to browse and download models by clicking	LM Studio
Want to script it, or serve models to other apps	Ollama
Want an OpenAI-compatible endpoint on localhost	Ollama
Are embedding inference into your own software	llama.cpp
Need the newest quantization formats or flags on day one	llama.cpp
Are shipping a commercial product	llama.cpp or Ollama (MIT)
Are on a Mac and want the fastest path	Either wrapper — both use MLX too

The thing that actually determines your experience

Not the runner. The model, and whether it fits in memory.

All three will run the same quantized model file at broadly similar speed, because the arithmetic is done by the same family of code. Not identical code, mind: Ollama pins its own llama.cpp version while LM Studio tracks another, and on Apple Silicon either may use MLX instead — so benchmarks do differ between them. They differ by margins that will not rescue a badly-chosen model.

What determines whether local AI feels good or useless is picking a model that fits your RAM with headroom left for the context window.

Switching runner to fix a slow or stupid model is solving the wrong problem. Get the model choice right first; the runner is a preference.

Our take

Start with LM Studio if you have never run a model locally — the graphical model browser removes an entire category of confusion, and you can move on later.

Move to Ollama the moment you want to automate anything. It is MIT, it serves an HTTP API, and it is the path of least resistance for scripts and other apps.

Reach for llama.cpp directly when you need control the wrappers do not expose, or when you are embedding inference in something you ship.

Most people should not think about this for more than five minutes. The runner is the least important decision in local AI.

What we addedThe layer diagram showing all three are the same engine (verified from Ollama's pinned LLAMA_CPP_VERSION and LM Studio's own docs), a licensing comparison that matters for commercial use, and a decision table based on interface rather than benchmarks.

This article was researched and drafted with AI assistance from the sources listed below, then checked and edited by Fiqhro Dedhen before publication. How we work.