# Running a local language model

## Hardware Recommendations

1. Large Language Models (LLMs) require a lot of computational power and therefore lots of electricity and dedicated hardware. This following advise assumes a small scale project and availability of consumer hardware.
2. Smaller models typically respond qualitatively worse than bigger ones, but they are faster, need less memory and might already be sufficient for your use case (so start out with the small ones and if need be, scale up).
3. The size of a model can be measured in number of parameters in its neural network. The "b" in the model name typically stands for **b**illion parameters. It also can be measured in terms of gigabytes required to load the model into your devices RAM/VRAM.
4. The model should always completely fit into VRAM (fast), otherwise layers will typically be offloaded to RAM (very slow) and if it doesn't fit in there either, it will use your harddrive, typically a SSD or HDD (abysmally slow).
5. The Hardware recommendation to maximize prompt processing and token generation speed is a device with high *bandwidth*. To date, modern GPU with lots of VRAM will satisfy this requirement best.

## High-level explanation

You can use any program that creates a server with OpenAI-compatible API.

After you started your service, you can do this:

1. The "Chat Model" field in AI preferences is editable, so you can enter any model you have downloaded
2. There is a field called "API base URL" in "Expert Settings" where you need to provide the address of an OpenAI-compatible API server

Voilà! You can use a local LLM right away in JabRef.

## Step-by-step guide for `ollama`

The following steps guide you on how to use `ollama` to download and run local LLMs.

1. Install `ollama` from [their website](https://ollama.com/download)
2. Select a model that you want to run. `ollama` provides [a large list of models](https://ollama.com/library) to choose from. Some popular models are for instance [qwen3:30b-a3b](https://ollama.com/library/qwen3), [`granite3.1-moe:3b`](https://ollama.com/library/granite3.1-moe), [`devkit/L1-Qwen-1.5B-Max`](https://ollama.com/devkit/L1-Qwen-1.5B-Max), [`mistral:7b`](https://ollama.com/library/mistral) or [`mistral-small3.1:24b`](https://ollama.com/library/mistral-small3.1).
3. When you have selected your model, type `ollama pull <MODEL>:<PARAMETERS>` in your terminal. `<MODEL>` refers to the model name like `gemma2` or `mistral`, and `<PARAMETERS>` refers to parameters count like `2b` or `9b`.
4. `ollama` will download the model for you
5. After that, you can run ollama serve to start a local web server. This server will accept requests and respond with LLM output. Note: The ollama server may already be running, so do not be alarmed by a cannot bind error. If it is not yet running, use the following command: `ollama run <MODEL>:<PARAMETERS>`
6. Go to JabRef Preferences -> AI
7. Set the "AI provider" to "OpenAI"
8. Set the "Chat Model" to the model you have downloaded in the format `<MODEL>:<PARAMETERS>`
9. Set the "API base URL" in "Expert Settings" to `http://localhost:11434/v1/`

Now, you are all set and can chat "locally".