Ollama, LM Studio, llama.cpp or vLLM: which app?

All of these load the same models. They differ in how you talk to them and what they are built for. The apps page lists every one with the formats it reads and the hardware it supports.

Point and clickDesktop appsLM Studio and Jan search, download and chat in one window. Best first choice for most people.

One commandCommand-line toolsOllama is the easy one: ollama run model. llama.cpp is the engine many others are built on, with the most options.

Many usersServersvLLM and SGLang batch many requests at once for teams and products. They need the whole model in GPU memory.

Choosing

Just want to chat: LM Studio or Jan.
Want to script it or use it from other tools: Ollama (local API).
On a Mac: LM Studio for MLX models, Ollama for GGUF (Macs).
On a phone, tablet or other mobile device: PocketPal AI on iPhone, iPad and Android, or ChatterUI on Android. You download one model file inside the app instead of running a command (mobile devices).
Want every setting, or the newest features first: llama.cpp directly.
Stories and role-play, or no install at all: KoboldCpp (one file), with SillyTavern in front if you like; llamafile packs a model into a single executable.
A model too big for one Mac: exo spreads it across several.
Serving several users from NVIDIA GPUs: vLLM or SGLang.
Coding: an agent such as OpenCode, Aider, Cline or Continue in front of any of the above (guide).
Chatting with your files: Open WebUI, AnythingLLM or GPT4All’s LocalDocs (guide).
Images: ComfyUI, or SwarmUI on top of it; Invoke for canvas editing; Fooocus for simple SDXL; Draw Things or DiffusionBee on a Mac (guide).

Which file type does each read?

The app decides the format: GGUF for most, MLX for Apple silicon, safetensors and its quantized variants for servers (formats). Each model page names the best app for your hardware, so you rarely need to choose from scratch.

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.