The .nbx format is NeuroBrix's universal model container — a ZIP archive that encapsulates everything needed to run an AI model on any supported hardware.
model.nbx
├── manifest.json # Model identity, family, generation type
├── topology.json # Execution flow, component connections, loops
├── runtime/
│ ├── defaults.json # User-tunable defaults (steps, guidance, temperature)
│ └── variables.json # Variable resolution contracts
├── components/<name>/
│ ├── graph.json # TensorDAG of ATen operations
│ ├── runtime.json # Component runtime config
│ ├── profile.json # Hardware profile hints
│ ├── config.json # Component configuration
│ ├── weights_index.json # Weight shard mapping
│ └── weights/
│ ├── shard_000.safetensors
│ └── shard_001.safetensors
└── modules/ # Scheduler, tokenizer, etc.
| Principle | Description |
|---|---|
| Model-independent engine | The graph and weights travel in the container; the engine needs no per-model code |
| Self-describing | Everything needed to run is inside the container |
| Hardware-agnostic | The same .nbx runs on every hardware target the engine executes on (see Install by hardware) |
| Transport-oriented | Designed for efficient distribution and caching |
Contains model identity, family classification, and generation type (image, text, audio, etc.).
Defines the execution flow — how components connect, loop structures (for diffusion steps), and data routing.
The TensorDAG — a directed acyclic graph of ATen-level operations that defines the mathematical computation.
Model weights stored in safetensors format, sharded into ~2GB chunks for efficient transfer and memory mapping.
When imported, .nbx files are extracted to ~/.neurobrix/cache/. The runtime never reads .nbx archives directly — it always works from the extracted cache.
~/.neurobrix/
├── store/ # Downloaded .nbx archives
└── cache/ # Extracted model files (runtime reads from here)
The .nbx containers are produced by the build toolchain. You download them from the Model Hub; you do not need to build them to use NeuroBrix.
neurobrix validate and neurobrix inspectNext: Runtime Architecture