Installing Ollama

Installing Ollama on Windows, macOS or Linux, downloading the models the workbench uses, and what your computer needs to run them.

Ollama is a free program that runs AI models on your own computer. The desktop workbench uses it for everything a model does: summaries, answers from your papers, claim checks, redrawing figures, writing code, and searching by meaning. Nothing leaves your computer, there is no key or account, and once a model is downloaded it works offline.

Ollama is optional. Without it, the workbench still keeps, reads and searches your papers by their words, and you can use a hosted model with your own key instead (see Models and keys). The cloud workbench cannot use Ollama.

1. Install Ollama

Download it from ollama.com/download.

To check that it is running, open http://127.0.0.1:11434 in a browser. It should say Ollama is running.

2. Download the models

Open a terminal (PowerShell on Windows, Terminal on a Mac) and run one command per model. Start with the first two, about 5 GB together:

You don't need to remember these. When a model the workbench wants is missing, Settings shows its command with a Copy button.

ollama list shows the models you have, and ollama rm <name> removes one.

3. Choose them in the workbench

1. Settings → Searching: under Searching by meaning, the model is marked installed. Choose it and press Use this model. Your papers are indexed while the workbench is idle, and searches use the model once that is done.
2. Settings → Writing: under The model that writes, choose ollama:llama3.1, or any model you have pulled, written as ollama: followed by its name. Under Which model does what, give a job its own model if you like: ollama:llava for figures, ollama:qwen2.5-coder for code.
3. If you started Ollama after the workbench, press Check again on the Searching tab.

What your computer needs

• Memory: a model like llama3.1 needs about 8 GB free while it works. 16 GB in the computer is comfortable.
• A graphics card Ollama can use (NVIDIA, AMD, or any Apple silicon Mac) makes it many times faster.
• Without one it still works, on the processor, but slowly. On a laptop without such a card, an answer from eight passages took about four minutes, and checking one claim between one and five minutes. Searching by meaning stays quick, and its indexing runs in the background.

The workbench gives a model up to ten minutes to answer. It sizes each request to fit the model, and refuses a request that is too long, saying why, rather than letting it be cut short without notice.

When it does not work