Ollama is an open source program for downloading and running open weights AI models on your computer with one command, like "ollama run" plus the model name. It also offers a local API so other apps can use those models. It has cloud models too, so check which one you pick.
A question from the app. Pick an answer.
What is Ollama for?
With something like "ollama run" and the model name, you are chatting. It runs on your computer and also offers cloud models.
That is what frameworks like PyTorch or MLX are for. Ollama is made for using models that come already trained.
That is Hugging Face's thing. Ollama has its own model library, but its strength is making them easy to run.
You install it on macOS, Windows or Linux and type "ollama run" with the name of a model from its library. The first time it downloads it; after that you can chat in the terminal or in its app, offline if the model is local.
While it runs, Ollama exposes an API on your own machine (by default at localhost:11434). That way an editor, an agent or your own code can ask it for answers as if it were a cloud service. With a local model, the computing stays on your computer too; with a cloud model, the request goes to the cloud.
Memory, above all: the model has to fit in RAM or in the graphics card's memory, like any local model. That is why many versions in its library come quantized, with fewer bits per weight. Pick a size that fits your computer and test whether the quality is enough.
Hugging Face is where a huge number of models are published; Ollama is the program that makes them easy to run. They work together: with "ollama run hf.co/" followed by the user and repository you can run a GGUF model published there.
In a web chat like ChatGPT, the model runs on the company's servers. With Ollama and a local open weights model, you download the weights, swap the model whenever you like, and your computer does the computing.
"With Ollama everything stays on my computer." Only if you use local models. Cloud models (marked as cloud) send your request to Ollama's servers and ask you to sign in. Check which one you pick, or turn off the cloud if you want everything local.
"Ollama is a model." No: it is the program that runs models made by others, like Llama, Gemma or Qwen. "It is for training." It is made for using models that are already trained; for training there are frameworks like PyTorch or MLX.