Requirements and choosing a model¶
What the server needs (operating system, memory, disk, GPU) and how to pick the language model that fits your hardware. See also Models and hardware.
Jarvis Automation is a pre-release (release candidate rc18.17). Builds are on the releases page; each one is marked Pre-release, so open the newest and look under Assets.
Choosing a model¶
The language model you run sets how quick and how accurate Jarvis feels, so pick it by your machine's memory:
- Mac with 16 GB: Qwen 3 4B (Q4_K_M).
- Mac with 24 GB or more: Qwen 3.5 9B (Q4_K_M, vision).
- PC or Linux box with a GPU: the largest model the setup wizard marks Fits.
- No usable GPU: Qwen 3 4B (Q4_K_M).
The setup wizard suggests one and you can change it later. Models and hardware has the full catalog, the engines and the reasons; Benchmarks has the numbers.
The server¶
The server runs on a machine that stays on. It is one program, jarvisd, with its database built in: there is no Docker to install.
| System | Processor | Notes |
|---|---|---|
| macOS | Apple silicon (M1 or later) | Intel Macs are not supported. Models run on the GPU through Metal. |
| Windows | 64-bit Intel or AMD (amd64) |
Runs as a Windows service. |
| Linux | 64-bit Intel or AMD (amd64), or ARM (arm64) |
Needs systemd. Models run on an NVIDIA (CUDA) or AMD (ROCm, Vulkan) GPU, or on the CPU. |
Each install page lists what it needs: macOS, Windows, Linux. Since rc18.4 there is also a container, for Unraid or Docker; releases up to rc18 have no image.
Memory¶
The language model takes most of the memory, so the model you choose sets how much you need:
- A Mac shares its memory with the GPU: 16 GB runs Qwen 3 4B, 24 GB or more runs Qwen 3.5 9B.
- A PC or Linux box with a GPU: the model has to fit the card's memory, with room for its context. Qwen 3.8 27B, the largest GGUF model, needs a 24 GB card.
- On the CPU, the model runs in system RAM. Qwen 3 4B is the one to use, and it is slow.
The setup wizard reads your hardware and marks each model Fits or not, so you don't have to work this out by hand.
Disk¶
Nothing large is downloaded until you pick the models in the setup wizard. Plan for the models you pick plus a few GB for the program, the engines it downloads with them and the database:
| Set of models | Download |
|---|---|
| Qwen 3 4B, Whisper small.en, the voice, speaker recognition and embeddings | about 3.5 GB |
| Qwen 3.5 9B (with its vision projector) and the same others | about 7.5 GB |
| Qwen 3.8 27B (with its vision projector), Whisper large-v3-turbo and the same others | about 19.5 GB |
The sizes of every model are in the catalog. They are stored in the server's data directory.