Skip to content

Ubuntu desktop and CLI

Spum 0.2.4 includes a native GTK desktop application and the inference engine for Ubuntu x86_64. The window talks to its bundled local engine; it does not use a browser or a remote inference service.

Install

Tested on Ubuntu 22.04 x86_64. The package needs GTK 3, libcurl, OpenSSL 3 and FFmpeg; installing with apt resolves these dependencies.

curl -fL https://spum.si/spum/releases/spum_0.2.4_amd64.deb -o spum_0.2.4_amd64.deb
sudo apt install ./spum_0.2.4_amd64.deb
spum --version
spum-desktop

You can also open Spum from your Applications menu. The installer places the desktop and CLI together under /usr/lib/spum, with command launchers under /usr/bin. Weights are downloaded separately.

Download Ubuntu package · Portable archive · Release checksums

First conversation

  1. Obtain a compatible GGUF model, then choose Import GGUF. Files are copied into ~/.spum/models; an existing file with the same name is not overwritten.
  2. Select Refresh, choose the model and select Load.
  3. Choose NVIDIA GPU before loading to request CUDA, or clear it to use the CPU. The status line reports the backend actually loaded.
  4. Enter a prompt and select Send. The number beside Send sets the output token limit.

The app retains its conversation locally in ~/.local/share/spum/conversation.json by default (or the configured XDG data directory). New chat clears the current conversation. Choose Light / dark to change the current window's theme.

NVIDIA acceleration

Install the appropriate NVIDIA driver for your GPU through your system administrator or Ubuntu's driver tools. The CUDA toolkit is not required for this packaged engine. If the driver or architecture is unavailable, the engine reports the issue and falls back to CPU.

spum run model.gguf --gpu --temperature 0 -p "Explain the measurement."
spum bench model.gguf --gpu --n-prompt 256 -n 128 --reps 5

CUDA was functionally checked on an H200. Supported CUDA families include Llama, Qwen 2/2.5/3 and MedGemma text inference. Support varies by architecture and quantization. Gemma 4 is not supported on the CUDA backend. Image decoding and image input are currently macOS-only; the Ubuntu interface does not offer image attachment. A GPU checkbox does not imply every model is supported on that GPU.

Local speech

Select Get speech models once to download and verify the English recognition and synthesis packs. Use Transcribe audio to choose an input and transcript destination. To synthesize speech, enter text in the composer, choose Speak text, and choose a WAV destination.

The same operations are available in the CLI:

spum models download whisper-base-en
spum models download kokoro-en-v1-int8
spum models verify whisper-base-en
spum models verify kokoro-en-v1-int8
spum transcribe recording.m4a --json --output transcript.json
spum speak --text "The scanner is ready." --voice af_heart --output speech.wav

See the speech guide for all options, supported English voices, pronunciation dictionaries and caption timing. Linux audio decoding uses the system FFmpeg package. The app does not record from a microphone or automatically play generated audio.

Portable archive

sudo apt install libgtk-3-0 libcurl4 libssl3 ffmpeg
mkdir spum-0.2.4
curl -fL https://spum.si/spum/releases/spum-0.2.4-linux-x86_64.tar.gz -o spum-linux.tar.gz
tar -xzf spum-linux.tar.gz -C spum-0.2.4
cd spum-0.2.4
./spum-desktop
./spum --help

Keep spum, spum-desktop and spum-runtime together. The archive includes notices, licences and the required public dependency source archive. No Python runtime is required.

Checks performed

The .deb was installed in Ubuntu 22.04, and its installed CLI, model catalog and native window startup were checked. The final portable engine was checked on Ubuntu 22.04 on the DGX: CPU and CUDA generation, the app's local API, missing-header rejection, synthesis and recognition. Both speech model packs were downloaded from the public Hugging Face repo and verified. These functional checks are not a new throughput benchmark or a clinical validation.

Updates and removal

Install the next .deb with sudo apt install ./NEW_PACKAGE.deb. To remove the system application:

sudo apt remove spum

Your downloaded models and local conversation data are not removed by that command. Inspect your own data before deleting it.

iPhone

There is no released or physically tested iPhone app in this release. Local inference on Apple silicon Macs and Ubuntu instrument computers is edge inference; it should not be represented as an iPhone performance result.

A Neuroscape Imaging product.