NeuroBrix is a universal deep learning inference engine. One runtime, one container format, one command set across every modality. The current stable engine is 0.5.4.
Benchmarks: measured speed, the limits, and the method. Figures live there and nowhere else.
| If you want to... | Go to |
|---|---|
| Get a first result in an hour | First hour |
| Pick the right install for your machine | Install by hardware |
| Browse and import models | Model Hub & Importing |
| Look up a command or flag | CLI Reference |
| Fix an error message | Troubleshooting |
The four verbs below drive every model in the same way:
neurobrix import <org>/<Model> # fetch a .nbx container from the hub
neurobrix serve --model <Model> # load it once into a background daemon
neurobrix run --prompt "..." # inference against the loaded model
neurobrix stop # release the memory
| One format | Every model ships as a single .nbx container carrying its own execution topology, weight map and runtime configuration. See The .nbx Format. |
| One solver | Prism reads the safetensors headers, sizes the job against the hardware actually present, and picks a placement strategy. See Runtime Architecture. |
| One registry | The hub serves every container over the same API the CLI uses, so neurobrix hub and neurobrix.es/catalog never disagree. See Model Hub. |
| Zero silent fallback | When an input is missing or a strategy will not fit, NeuroBrix refuses and says why instead of quietly degrading the output. See ZERO Principles. |
Every row below is published and downloadable now. This section is generated from the hub; it is not maintained by hand.
49 models, 10 categories, 9 runtime families, 0.96 TiB packaged.
Generated from the hub on 2026-11-09 — do not edit by hand, runnode scripts/wiki-sync.mjs catalogue.
Live view: neurobrix.es/catalog · API: /api/models
LLM (4)| Model | Size | Licence | |
|---|---|---|---|
TinyLlama/TinyLlama-1.1B-Chat-v1.0 |
2.1 GiB | apache-2.0 | |
Qwen/Qwen3-30B-A3B-Thinking-2507 |
56.9 GiB | apache-2.0 | |
deepseek-ai/deepseek-moe-16b-chat |
30.5 GiB | deepseek | |
ibm-granite/granite-3.1-1b-a400m-instruct |
2.6 GiB | apache-2.0 |
neurobrix hub --category LLM
CODE (2)| Model | Size | Licence | |
|---|---|---|---|
deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct |
29.9 GiB | deepseek | |
Qwen/Qwen3-Coder-30B-A3B-Instruct |
56.9 GiB | apache-2.0 |
neurobrix hub --category CODE
VLM (1)| Model | Size | Licence | |
|---|---|---|---|
zai-org/GLM-4.1V-9B-Thinking |
19.2 GiB | mit |
neurobrix hub --category VLM
MULTIMODAL (5)| Model | Size | Licence | |
|---|---|---|---|
Qwen/Qwen3-VL-30B-A3B-Thinking |
57.9 GiB | apache-2.0 | |
deepseek-ai/Janus-Pro-7B |
13.8 GiB | mit | |
Qwen/Qwen3-Omni-30B-A3B-Instruct |
65.8 GiB | apache-2.0 | |
openbmb/MiniCPM-o-4_5 |
19.3 GiB | apache-2.0 | |
inclusionAI/Ming-Lite-Omni-1.5 |
51.1 GiB | mit |
neurobrix hub --category MULTIMODAL
TTS (5)| Model | Size | Licence | |
|---|---|---|---|
ResembleAI/chatterbox |
2.0 GiB | mit | |
hexgrad/Kokoro-82M |
353 MiB | apache-2.0 | |
fishaudio/s1-mini |
2.6 GiB | cc-by-nc-sa-4.0 | |
canopylabs/orpheus-3b-0.1-ft |
14.1 GiB | apache-2.0 | |
microsoft/VibeVoice-1.5B |
5.0 GiB | mit |
neurobrix hub --category TTS
STT (3)| Model | Size | Licence | |
|---|---|---|---|
openai/whisper-large |
5.8 GiB | mit | |
openai/whisper-large-v3-turbo |
1.5 GiB | mit | |
nvidia/parakeet-tdt-1.1b |
4.0 GiB | cc-by-4.0 |
neurobrix hub --category STT
AUDIO_LLM (3)| Model | Size | Licence | |
|---|---|---|---|
ibm-granite/granite-speech-3.3-8b |
16.1 GiB | apache-2.0 | |
nvidia/canary-qwen-2.5b |
4.7 GiB | cc-by-4.0 | |
mistralai/Voxtral-Mini-3B-2507 |
8.7 GiB | apache-2.0 |
neurobrix hub --category AUDIO_LLM
IMAGE (5)| Model | Size | Licence | |
|---|---|---|---|
Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers |
9.0 GiB | nvidia-open-model-license | |
PixArt-alpha/PixArt-Sigma-XL-2-1024-MS |
20.3 GiB | openrail++ | |
PixArt-alpha/PixArt-XL-2-1024-MS |
20.3 GiB | openrail++ | |
ostris/Flex.1-alpha |
24.5 GiB | apache-2.0 | |
Efficient-Large-Model/Sana_1600M_4Kpx_BF16_diffusers |
12.0 GiB | nvidia-open-model-license |
neurobrix hub --category IMAGE
UPSCALER (10)| Model | Size | Licence | |
|---|---|---|---|
JingyunLiang/SwinIR-001_classicalSR_DIV2K_s48w8_SwinIR-M_x4 |
43 MiB | apache-2.0 | |
caidas/swin2SR-classical-sr-x2-64 |
47 MiB | apache-2.0 | |
xinntao/Real-ESRGAN-RealESRGAN_x4plus |
64 MiB | bsd-3-clause | |
XPixelGroup/HAT-HAT-L_SRx4_ImageNet-pretrain |
146 MiB | apache-2.0 | |
caidas/swin2SR-classical-sr-x4-64 |
48 MiB | apache-2.0 | |
caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr |
47 MiB | apache-2.0 | |
JingyunLiang/SwinIR-001_classicalSR_DF2K_s64w8_SwinIR-M_x2 |
43 MiB | apache-2.0 | |
ai-forever/Real-ESRGAN-RealESRGAN_x2 |
64 MiB | — | |
XPixelGroup/HAT-HAT-S_SRx4 |
35 MiB | apache-2.0 | |
ai-forever/Real-ESRGAN-RealESRGAN_x8 |
64 MiB | — |
neurobrix hub --category UPSCALER
VIDEO (11)| Model | Size | Licence | |
|---|---|---|---|
Wan-AI/Wan2.1-T2V-1.3B-Diffusers |
26.9 GiB | apache-2.0 | |
zai-org/CogVideoX-2b |
12.9 GiB | apache-2.0 | |
Wan-AI/Wan2.2-I2V-A14B-Diffusers |
118.0 GiB | apache-2.0 | |
Wan-AI/Wan2.1-I2V-14B-480P-Diffusers |
84.4 GiB | apache-2.0 | |
hpcai-tech/Open-Sora-v2 |
42.4 GiB | apache-2.0 | |
rhymes-ai/Allegro |
23.6 GiB | apache-2.0 | |
rhymes-ai/Allegro-TI2V |
24.3 GiB | apache-2.0 | |
genmo/mochi-1-preview |
38.2 GiB | apache-2.0 | |
Wan-AI/Wan2.1-VACE-1.3B-diffusers |
18.2 GiB | apache-2.0 | |
zai-org/CogVideoX-5b-I2V |
21.0 GiB | cogvideox-license | |
Efficient-Large-Model/SANA-Video_2B_720p_diffusers |
17.1 GiB | nvidia-open-model-license |
neurobrix hub --category VIDEO
| Section | Description |
|---|---|
| First hour | From an empty machine to a first result |
| Install by hardware | The right install for NVIDIA, Apple Silicon or CPU |
| Installation | Install the engine, and what it puts on your disk |
| Quickstart | Get to a first result |
| Licences | Gated models and licence acceptance |
| Model Hub & Importing | Browse, import, manage local models |
| Inference Guide | Run each kind of model |
| Serving Guide | Keep a model loaded between runs |
| CLI Reference | Every command and flag |
| API Reference | The hub's HTTP API |
| Troubleshooting | Error messages and their fixes |
| Changelog | What changed in 0.5.3 and 0.5.4 |
| The .nbx Format | The universal model container |
| Runtime Architecture | How the engine executes a model |
| Hardware Support | Detection, what executes today, strategies |
| Hardware Profiles | YAML profile reference |
| Benchmarks | Measured speed, the limits, and the method |
| Roadmap | What has shipped and what is open |
| Python | 3.10, 3.11 or 3.12 |
| PyTorch | required; install it first for your hardware, see Install by hardware |
| Disk | the engine, plus the models themselves; the catalogue shows each size. Import needs free space of about twice the archive plus 2 GiB |
| GPU | NVIDIA with CUDA for anything substantial. Apple Silicon (MPS) and CPU also execute |
Stated here rather than discovered after the download.
sha256 for every container; the CLI checks size and archive integrity only.Next: First hour
© 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