Verified against neurobrix 0.6.2 (the --help of every command at tag v0.6.2, compared with v0.5.4). Run any command with --help for the authoritative list.
neurobrix --version # NeuroBrix v0.6.2
A subcommand is required. Commands: run, serve, chat, stop, upscale, hub, import, list, remove, clean, info, inspect, validate, doctor.
calibrate, autotune, drift and coverage are present in 0.6.2: present, not documented.
| Code | Meaning |
|---|---|
| 0 | Success, --help and --version (doctor: no blocking problem found) |
| 1 | ERROR: ... (missing file, invalid value) or UNEXPECTED ERROR: ... with a traceback; doctor: problems found |
| 2 | No subcommand or an unknown flag; hub: the registry rejected the request (for example an unknown category) |
| 130 | Interrupted with Ctrl+C |
run exits with the non-zero status of the run itself. The exit code of validate is not documented yet.
--json (on info, inspect, list, remove, import, hub, doctor) prints one JSON record on stdout and the human-readable lines on stderr. import --json emits one JSON event per line.
import takes the full reference org/Model (case-sensitive). Once installed, a model is known by its bare name: use that with run, serve, upscale, inspect, remove.
neurobrix import hexgrad/Kokoro-82M
neurobrix run --model Kokoro-82M --prompt "Hello" --output hello.wav
neurobrix hubBrowse the registry (no sign-in needed).
| Flag | Description |
|---|---|
--category, -c |
Filter by category (upper-cased for you) |
--search, -s |
Search by text |
--registry |
Registry URL (default https://neurobrix.es) |
--json |
Machine-readable output |
neurobrix hub --category IMAGE
An unknown category prints ERROR: ... followed by Valid categories: ... (the list comes from the registry) and exits 2. The valid values are in Model Hub.
neurobrix import ORG/NAMEDownload and install a model.
| Flag | Description |
|---|---|
--registry |
Registry URL |
--force |
Reinstall a model that is already installed |
--no-keep |
Delete the downloaded .nbx after extraction (default keeps it in ~/.neurobrix/store/) |
--accept-license |
Accept a gated licence without the prompt |
--json |
One JSON event per line |
neurobrix import hexgrad/Kokoro-82M
What happens, in order: the model is looked up (Model 'org/name' not found on registry. if absent), the licence is checked, the download is requested and streamed, the archive is checked, then extracted to ~/.neurobrix/cache/<name>.
<file>.nbx.part. Re-run the same command and it resumes with an HTTP Range request.ERROR: not enough disk for this import). Peak use is about twice the archive size unless you pass --no-keep.Model already installed and exits 0 unless --force.Accept license terms? [yes/No]. In a non-interactive shell pass --accept-license or set NBX_ACCEPT_LICENSE=1, otherwise: ERROR: license '<id>' requires explicit acceptance: re-run with --accept-license (exit 1). Acceptance is remembered locally. See Licences.sha256sum) against the sha256 in the API.| Command | Description |
|---|---|
neurobrix list [--store] [--json] |
Installed models (--store: also the downloaded archives) |
neurobrix remove MODEL [--store] [--all] [--json] |
Remove a model. Default removes the extracted copy only; --store the archive, --all both |
neurobrix clean [--store] [--cache] [--all] [-y] |
Empty the store and/or the cache. With none of the three flags: ERROR: Specify --store, --cache, or --all (exit 1) |
| Command | Description |
|---|---|
neurobrix info [--models] [--hardware] [--system] [--json] |
Engine, models and hardware summary |
neurobrix inspect PATH_OR_MODEL [--topology] [--weights] [--json] |
Look inside a .nbx file or an installed model |
neurobrix validate NBX... [--level LEVEL] [--strict] [--json] [-v] |
Check one or more .nbx files. Levels: structure, schema, coherence (default), deep |
neurobrix doctor [--json] |
Check the install: PATH, the Triton engine (nbxtensor and Triton versions), PyTorch and its CUDA build (optional), GPUs and compute capability, GPU drivers |
doctor exits 0 with "No blocking problem found", or 1 with the problems and the fix for each (1 also when no engine can run). PyTorch engine: not installed (optional) is a blocking problem only when the Triton engine cannot run either; the PyTorch wheel it suggests depends on the GPU (a V100 gets cu126). With --json it prints {"schema", "engine", "ok": bool, "problems": [...]}. It checks NVIDIA GPUs only; it does not check ROCm or Apple GPUs. See Install by hardware.
neurobrix runneurobrix run --model <name> --prompt "..." [options]
--model may be omitted only when a serving daemon is running (ERROR: --model is required when no daemon is running.). While a daemon is up, a cold run is refused (ERROR: A serving daemon is already running (PID n).); with a matching model the run uses the warm daemon.
Common options
| Flag | Description |
|---|---|
--hardware |
Hardware profile (auto-detected if omitted) |
--explain-plan |
Show the execution plan |
--json |
Machine-readable result |
--prompt |
Text prompt |
--output |
Output file. Default output_<model>.<ext>; the extension must match the model family |
--seed, --set KEY=VALUE |
Seed; override a model default (repeatable) |
--certified-only |
Same as NBX_AUTOTUNE_CERTIFIED_ONLY=1 (also on calibrate and drift) |
--mode |
auto, text, chat, agent, image, audio, video, t2i, img2img, inpainting, t2v, i2v, v2v |
Execution mode (one at most; the default is --compiled when PyTorch is installed, --triton otherwise): --compiled, --sequential, --triton, --triton-sequential. Passing more than one: ERROR: Only one execution mode flag can be passed at a time.
Per family
| Flag | Used by |
|---|---|
--max-tokens, --temperature, --top-k, --top-p, --repetition-penalty, --system, --chat / --no-chat |
Language models |
--steps, --cfg, --height, --width |
Image and video generation |
--input-image, --mask-image, --reference-image |
Image-to-image, inpainting, vision-language |
--num-frames, --fps, --input-video, --video |
Video |
--audio |
Speech-to-text, audio language models |
--reference-audio, --speaker |
Text-to-speech |
--workdir, --max-turns, --bash-timeout, --allow-network, --approve-all, --yolo, --transcript-dir |
Agent mode |
A first command for each family
neurobrix run --model <llm> --prompt "Explain ZFS in one paragraph" --max-tokens 256
neurobrix run --model <image> --prompt "a lighthouse at dusk" --steps 30 --output out.png
neurobrix run --model <video> --prompt "waves on a beach" --num-frames 33 --output out.mp4
neurobrix run --model <tts> --prompt "Hello" --output hello.wav
neurobrix run --model <stt> --audio speech.wav
neurobrix run --model <vlm> --prompt "Describe this" --input-image photo.png
Required inputs depend on the family. A missing one stops the run with ZERO FALLBACK: family '<f>' requires <flag>. Provide it on the command line., and a wrong output extension with ZERO FALLBACK: output extension '.x' incompatible with family .... Vision-language models need --prompt and --input-image; audio language models need --audio and --prompt; multimodal models need --mode text|image|audio.
Where speech-to-text writes its transcript when --output is omitted is not documented yet.
neurobrix upscaleneurobrix upscale --model <name> --input in.png --output out.png
| Flag | Description |
|---|---|
--model |
Required |
--input, --input-image |
Required, the source image |
--output |
Required, a PNG path |
--mode |
compiled (default), sequential, triton, triton-sequential |
--hardware |
Hardware profile |
A missing input prints ERROR: input image not found: <path> (exit 1).
neurobrix serve --model <name> # load once
neurobrix run --prompt "..." # no --model needed
neurobrix chat # interactive, language models
neurobrix stop # shut it down
| Command | Flags |
|---|---|
serve |
--model (required), --hardware, --timeout (idle seconds, default 1800), --foreground, --ready-fd / --ready-file (for a program that launches the daemon: where to write the one-line JSON ready record), execution-mode flags |
chat |
--max-tokens, --temperature, --repetition-penalty, --agent and the agent flags. In chat: /new, /context, /status, /quit |
stop |
none; asks the daemon to exit, then terminates it if needed, and cleans stale files |
If a daemon already exists, serve prints [Serve] Daemon already running (PID n) and exits 1. One task at a time. More in the Serving guide.
Next: Configuration
© 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