One engine. Any modality.
Any hardware.
One runtime, one .nbx format, one command set — and 49 models across 10 categories running on it today: language and code, vision, speech in and out, audio-native LLMs, images, upscalers, video, and omni models that do several at once.
- 49
- models
- 10
- categories
- 9
- runtime families
- 27
- publishers
- 984.5 GiB
- packaged
One format
Every model ships as a single .nbx container carrying its own execution topology, weight map and runtime configuration. Nothing to wire up per model.
One command set
import, serve, run, stop. The same four verbs drive every package in the catalogue, from the smallest upscaler to the largest video model. No per-model scripts, no bespoke runtimes.
One solver
Prism reads the safetensors headers, sizes the job against the GPUs actually present, and picks a placement strategy — from single-GPU to ZeRO-3 with CPU offload.
Zero silent fallback
If a model says it runs, it runs. When a strategy will not fit, NeuroBrix says so instead of quietly degrading the output.
The catalogue
Not a roadmap. Every model below is packaged, published and downloadable right now, and every count on this page is read from the hub as the page is served.
- TinyLlama/TinyLlama-1.1B-Chat-v1.0
- Qwen/Qwen3-30B-A3B-Thinking-2507
- deepseek-ai/deepseek-moe-16b-chat
- + 1 more
- deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct
- Qwen/Qwen3-Coder-30B-A3B-Instruct
- zai-org/GLM-4.1V-9B-Thinking
- Qwen/Qwen3-VL-30B-A3B-Thinking
- deepseek-ai/Janus-Pro-7B
- Qwen/Qwen3-Omni-30B-A3B-Instruct
- + 2 more
- ResembleAI/chatterbox
- hexgrad/Kokoro-82M
- fishaudio/s1-mini
- + 2 more
- openai/whisper-large
- openai/whisper-large-v3-turbo
- nvidia/parakeet-tdt-1.1b
- ibm-granite/granite-speech-3.3-8b
- nvidia/canary-qwen-2.5b
- mistralai/Voxtral-Mini-3B-2507
- Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers
- PixArt-alpha/PixArt-Sigma-XL-2-1024-MS
- PixArt-alpha/PixArt-XL-2-1024-MS
- + 2 more
- JingyunLiang/SwinIR-001_classicalSR_DIV2K_s48w8_SwinIR-M_x4
- caidas/swin2SR-classical-sr-x2-64
- xinntao/Real-ESRGAN-RealESRGAN_x4plus
- + 7 more
- Wan-AI/Wan2.1-T2V-1.3B-Diffusers
- zai-org/CogVideoX-2b
- Wan-AI/Wan2.2-I2V-A14B-Diffusers
- + 8 more
Triton engine faster than PyTorch in 30 of 36 pairs on V100
NeuroBrix 0.6.2, 18 models on both engines, medians of 3 runs, every figure from one dated register. The six pairs where it is slower are published too.
Zero to first generation
Four commands. The model below is the smallest language model in the catalogue (2.1 GiB), so this runs on a laptop before you commit a GPU to anything larger.
$ pip install neurobrix
$ neurobrix hub# the same 49 models as this page
$ neurobrix import TinyLlama/TinyLlama-1.1B-Chat-v1.0
$ neurobrix serve --model TinyLlama-1.1B-Chat-v1.0
$ neurobrix chat
What NeuroBrix does not do yet
Stated here rather than discovered after the download.
- NVIDIA / CUDA only. AMD (ROCm), Intel (oneAPI, Arc) and Apple (MPS) are detected but not yet executed on. CPU execution works for small models and is for development, not throughput.
- Python 3.10+ and PyTorch 2.0+ are required. There is no standalone binary; NeuroBrix installs into a Python environment.
- Models are large. The catalogue runs from 34.8 MiB to 118 GiB per package (Wan-AI/Wan2.2-I2V-A14B-Diffusers). Budget the disk and the VRAM before importing; the biggest video models want a multi-GPU host.
- Licences are the publisher's, not ours. Each package carries the upstream licence and some require accepting it before download. Check the model page before you build on it.