# System requirements

Axon runs its models on your own machine, so what decides whether it is
usable is the GPU. It picks the size of model that fits the machine; more GPU
memory buys a larger model and better answers.

## At a glance

| | Minimum | Recommended |
|---|---|---|
| **Windows / Linux** | A graphics card with **4 GB** of its own memory, 8 GB RAM | A card with 8 GB or more, 16 GB RAM |
| **Mac** | **Apple silicon**, 8 GB | Apple silicon, 16 GB or more |
| **Disk** | 10 GB free | More, for large document sets |

A machine without a GPU — or with only the graphics built into the processor,
including every Intel Mac — will install and open Axon, but answers come too
slowly to be practical.

## Operating system

| System | Supported |
|---|---|
| macOS | 10.15 Catalina or later, Apple silicon or Intel |
| Windows | Windows 10 or 11, 64-bit (x64) |
| Linux | 64-bit (x86-64), Ubuntu 24.04 or a distribution of the same age or newer |

On Windows, Axon uses Microsoft Edge WebView2. Windows 11 includes it; on
Windows 10 the installer fetches it if it is missing. The Linux `.deb`
package needs WebKitGTK 4.1 and GTK 3, which `apt` installs with it.

## Graphics

**A GPU is required for Axon to be usable.**

- **Windows and Linux** — a graphics card with at least **4 GB** of its own
  memory (NVIDIA or AMD). Axon uses it through Vulkan, which current drivers
  provide; keep the driver up to date.
- **Apple silicon** — the GPU is built in and used automatically.
- **Intel Macs, and PCs with only built-in graphics** — Axon runs on the
  processor, which is too slow for everyday use.

## The models Axon runs

Everything Axon does with your documents runs on these models, all on your
machine and all fetched by the Setup assistant on first launch:

| Job | Model | Size | When it runs |
|---|---|---|---|
| **Chat** — answers questions, runs skills | Qwen3.5, sized for your machine (below) | 0.5–5.7 GB | Whenever you ask something |
| **Summaries** — a summary of each document, written while indexing | The chat model, unless you pick another in **Settings → LLM → Models** | — | While indexing |
| **Embedding** — turns passages into what search matches against | BGE-large-en-v1.5 | 0.7 GB | While indexing, and on every question |
| **OCR / vision** — reads scanned PDF pages and images | PaddleOCR-VL 1.6 | 1.8 GB | Only while indexing scans and images; unloaded after two minutes idle |

The chat and embedding models stay loaded side by side. The vision model joins
them while you index scanned documents, so a large batch of scans is when the
GPU is busiest.

## Which chat model you get

The chat model is always Qwen3.5; Axon picks its size from the memory it will
run in, after leaving room for the embedding model.

**With a graphics card**, that is the card's free memory, not the computer's.
Part of it is always taken by the embedding model and working space, so in
practice:

| Graphics card | Chat model |
|---|---|
| 4 GB | Qwen3.5 2B |
| 6 GB or 8 GB | Qwen3.5 4B |
| 10 GB or more | Qwen3.5 9B |

**On Apple silicon**, it is the Mac's memory, which the GPU shares:

| Memory | Chat model |
|---|---|
| 8 GB | Qwen3.5 2B |
| 16 GB | Qwen3.5 4B |
| 32 GB or more | Qwen3.5 9B |

Close other GPU-heavy programs before a long session — what they hold is not
free for Axon. **Settings → LLM → On this machine** shows which model Axon
chose and the GPUs it uses.

## Disk space

| What | Size |
|---|---|
| The local models | 2.5 GB for embedding and OCR, plus the chat model — about 3.8 GB in all with the 2B, 8.2 GB with the 9B |
| Search indexes | Grows with the documents you add; plain text takes far less than its original files |

Allow **10 GB free** to install Axon and start working. Your documents are not
copied — a project indexes a folder where it already sits.

## Network

Axon needs the internet three times: to download it, to download the local
models on first launch, and to fetch updates when you ask for them. After that
it answers questions with no connection at all.

Some features use a network on purpose, and only when you turn them on:

- **Sharing a project** (Pro) uses your local network between the two
  computers.
- **Cloud-drive search** (Pro) and **web search** for connectors reach the
  service you connected.
- **Pro** connects to exaflux.io now and then to renew the licence.
- **Your own Ollama server**, if you point Axon at one.

See [Settings & providers](https://exaflux.io/docs/axon/settings-and-providers#what-runs-where) for
exactly what runs where.