Ollama vs LM Studio: which program for running AI models locally?

What this is about

A local AI model is a language model you download and run on your own computer. Our article on local AI models explains the benefits and limits. This page is only about choosing the program you use to run such models.

According to PC Games Hardware, LM Studio and Ollama are the two programs most often recommended for getting started. LM Studio is aimed at anyone who would rather not open a terminal, the text window for commands. It combines a ChatGPT-style chat window with a model search that connects directly to the Hugging Face platform. According to the magazine, Ollama takes the opposite approach: stripped down to the essentials, traditionally used from the command line, and now with a simple chat window.

Setup and hardware

Ollama: according to the project page, you install Ollama on macOS and Linux with one terminal command or by download, and on Windows likewise; there is also a Docker version. On Windows, Ollama says the installation does not need administrator rights. According to the documentation, it requires macOS 14 or newer, with Macs on M chips using the graphics unit and older Intel Macs only the main processor, and Windows 10 22H2 or newer. A graphics card is not required: according to the documentation, Ollama can also run models using only the main processor and normal RAM, just more slowly. If you want to use an Nvidia or AMD graphics card for acceleration, you need current drivers for it. On Windows, Ollama says the program itself takes at least 4 gigabytes, plus the models.

LM Studio: you download the app and do everything with the mouse. According to LM Studio, a Mac needs an Apple chip from M1 and macOS 14 or newer; Intel Macs are not supported. 16 gigabytes of RAM are recommended; with 8 gigabytes only small models work. On Windows, LM Studio runs on PCs with an x64 processor, which needs the AVX2 instruction set, and also on Windows PCs with an ARM chip such as the Snapdragon X Elite; and 16 gigabytes of RAM and at least 4 gigabytes of graphics memory are recommended.

Parameters are the internal numerical values a model is made of; more parameters usually mean a larger model. According to PC Games Hardware, 8 gigabytes of graphics memory can be enough for compact models up to about nine billion parameters and 16 gigabytes for models around 20 billion parameters, but only for memory-saving compressed versions, so-called quantised models. On top of that comes memory for the conversation history, which grows with the length of the conversation. So check the file size of the specific model version and allow some headroom.

Licences: program and model

According to the project page, Ollama is released under the MIT licence. This is a free open-source licence: you may use the program at work, modify it and pass it on.

LM Studio is free but not open source; the consultancy Qytera lists it as proprietary. Its terms allow use for personal purposes and for internal business purposes.

Important: the program’s licence is not the model’s licence. LM Studio points out in its terms that downloaded models come from third parties and that their own licence terms apply. If you want to use results commercially, check the licence of each model, whichever program you use.

Cloud offers: when you are no longer local

Both providers now earn money with cloud models that run on other people’s servers.

Ollama: according to its pricing page there is a free plan with a small starter credit, Pro at 20 dollars a month or 200 dollars a year with 60 dollars of usage credit per month, and Max at 100 dollars. The included credit renews monthly, even on annual billing. You do not need a subscription for more cloud use, though: even with a free account you can buy usage credit and unlock all cloud models with it. Ollama says the models run mainly in the US and in Europe when extra capacity is needed, and that prompts and responses are never logged or used for training. Local models remain free without limits.

LM Studio: the app is now called LM Studio Bionic and includes an agent for documents and code. According to its pricing page, running locally is free, Bionic+ with US-hosted open models costs 20 dollars and Pro 100 dollars a month. LM Studio says its cloud services do not store prompts and responses.

These are vendor claims. If privacy matters to you, make sure with either program that the model you choose really runs locally.

Which program for whom?

The following is an editorial recommendation based on the providers’ information and PC Games Hardware.

LM Studio suits you if you want to start without a terminal, like browsing for models with the mouse and have a Mac with an Apple chip or a Windows PC with enough memory.

Ollama suits you if you want an open, lean tool that also runs on Linux, in Docker or on older Intel Macs, or if you want to connect the AI to other programs later. According to Qytera, Ollama is the usual default for server setups.