neurobrix run handles every model family. What you must pass depends on the family.
| Mode | Commands | Best for |
|---|---|---|
| Warm (recommended) | neurobrix serve --model <name>, then neurobrix run --prompt "..." |
Several runs, interactive use |
| Cold | neurobrix run --model <name> --prompt "..." |
One-off runs, scripts |
A cold run loads the weights, runs and unloads. While a daemon is running, a cold run is refused with ERROR: A serving daemon is already running (PID n).; run against the warm daemon instead. See Serving Guide.
| Family | Command | Output |
|---|---|---|
| LLM | neurobrix run --model <name> --prompt "..." |
.txt |
| Image | neurobrix run --model <name> --prompt "..." --output x.png |
.png |
| Video | neurobrix run --model <name> --prompt "..." --output x.mp4 |
.mp4 |
| TTS | neurobrix run --model <name> --prompt "text" --output x.wav |
.wav |
| STT | neurobrix run --model <name> --audio f.wav |
text; where it goes without --output: not documented yet |
| Upscaler | neurobrix upscale --model <name> --input in.png --output out.png |
.png |
| VLM | --prompt and --input-image, both required |
text |
| Audio LLM | --audio and --prompt, both required |
text |
| Multimodal | --prompt and --mode text, image or audio; the mode is required |
per mode |
If a required input is missing, the engine stops with ZERO FALLBACK: family '<f>' requires <flag>. Provide it on the command line. A wrong --output extension is refused the same way, naming the expected one. Without --output, the file is named output_<model> plus the extension for the family (the --help text still says output.png).
| Flag | Applies to | Meaning |
|---|---|---|
--prompt |
most families | Input text |
--steps, --cfg, --seed |
image, video | Diffusion steps, guidance scale, random seed |
--height, --width |
image | Output size |
--num-frames, --fps |
video | Clip length and rate |
--input-image |
image, video, VLM, upscaler | Input picture; on image it selects img2img, on video i2v |
--mask-image |
image | Inpainting mask; needs --input-image |
--video |
video | Input clip, selects v2v |
--reference-audio, --speaker |
TTS | Voice reference and speaker id |
--max-tokens, --temperature, --repetition-penalty, --top-k, --top-p |
LLM | Sampling |
--system |
LLM | System prompt |
--chat / --no-chat |
LLM | Mutually exclusive |
--mode |
all | auto, text, chat, agent, image, audio, video, t2i, img2img, inpainting, t2v, i2v, v2v |
--set KEY=VALUE |
all | Override a model default; repeatable |
--hardware |
all | Hardware profile id |
--explain-plan |
all | Show the placement plan |
--json |
all | Machine-readable output |
Defaults for steps, guidance and size come from the model itself. Per-family defaults for sampling flags such as temperature: not documented yet.
Image:
neurobrix run --model <name> --prompt "A cute robot in a garden" --steps 20 --seed 42 --cfg 5.0 --width 1024 --height 1024 --output robot.png
Expected result: a PNG at robot.png.
Override a default:
neurobrix run --prompt "test" --set global.guidance_scale=7.5
Pass at most one on run or serve:
| Flag | Meaning |
|---|---|
--compiled |
Default |
--sequential |
Native dispatcher, for debugging |
--triton |
Triton kernels; Triton is not a declared dependency, and neurobrix doctor says whether it is available |
--triton-sequential |
Both |
More than one gives ERROR: Only one execution mode flag can be passed at a time. and exit code 1.
Next: Serving Guide
© 2025–2026 Hocine Benkelaya. NeuroBrix engine released under the Apache License 2.0.
Developed by Hocine Benkelaya · Supported and funded by Neural Network Holding Ltd · Commercial distribution by WizWorks OÜ
Documentation source: GitLab