NeuroBrix
Open-source engine that runs language, vision, speech, image and video models from one container format.
Tested on Linux
Use it
Linux x86_64
You need first
- Runtimes:
Python 3.10 or newer,PyTorch 2.1 or newer - A GPU is optional
uv tool install neurobrix
uv tool install neurobrix
Where the README and reality differ
- README: neurobrix run --model Kokoro-82M --prompt "Hello from NeuroBrix." --output speech.wav
What worked: add --speaker <voice> e.g. --speaker af_sky; without it Kokoro-82M aborts with ZERO FALLBACK listing 54 voices
Check that it works
neurobrix --versionThe output should contain “NeuroBrix v0.5.4”.
First thing to try
neurobrix doctorRun neurobrix doctor first to check your setup.
What it does
- Turn text into speechTested
How to use it
- Import a speech model with
neurobrix import hexgrad/Kokoro-82M --accept-license. - Run
neurobrix run --model Kokoro-82M --prompt "<your text>" --speaker af_sky --output <file>.wav. - Look for the SAVED line with your audio file name.
- Import a speech model with
- Generate images and videos, and upscale imagesTested
How to use it
- Import an upscaling model with
neurobrix import xinntao/Real-ESRGAN-RealESRGAN_x4plus --accept-license. - Run
neurobrix upscale --model Real-ESRGAN-RealESRGAN_x4plus --input <image> --output <output image>. - Look for the SAVED line with the upscaled file name.
- Import an upscaling model with
- Run language, vision, speech and video modelsNot tested
- Transcribe speech to textNot tested
How far pickfit got
Linux x86_64
- Installs
- Starts
- 2 of 4 features work
- Works offline
Good for
- transcribe audio to text with speech-to-text model
- synthesize speech audio from text
- generate an image from a text prompt
- generate video from a text prompt
- upscale an image with super-resolution model
How pickfit tested it
Each run starts from a clean machine and follows the steps above. How pickfit verifies
Linux x86_64
Its main function works
- Replayed from scratch
- Works offline
- Checked on 30 Oct 2026
- Version 0.5.4
Differences the test machine hit
- README: pip install torch matching hardware then pip install neurobrix
In the test: uv tool install neurobrix pulled torch 2.15.0+cu132 plus ~2.4 GB of NVIDIA CUDA wheels on a CPU-only box; doctor flags the waste and points to the cpu index build - README: neurobrix import <org>/<name> asks to accept the license
In the test: re-run with --accept-license (or NBX_ACCEPT_LICENSE=1) for non-interactive import