NeuroBrix is not one program. It is a set of separate products, each with its own repository, its own releases and one responsibility. Every function is filed in exactly one of them.
| Brick | What it owns | What it leaves to the others |
|---|---|---|
| NeuroBrix engine | The decisions: reading .nbx containers, Prism (memory and placement across your GPUs), the model families, the execution flow, the command line (neurobrix) and serving (neurobrix serve, an OpenAI-compatible API). |
It has no tensor code of its own: every tensor operation goes through nbxtensor. |
| nbxtensor | The tensor library: tensors and dtypes, Triton kernels, device memory and its allocator, the launcher, replay, page gather and certified tuning. Usable on its own, like torch for NeuroBrix. | It knows no model and has no command line. |
| NBXGrammar | Constrained decoding, in Rust: it compiles a JSON Schema, or a model's own tool-call format, into the token mask applied at every step. | It runs no model; the engine calls it while decoding. |
| NBX Code (to come) | The coding agent (command nibux): it will decide each action by your permissions, run it in a sandbox the operating system enforces and record it in a journal you can replay. |
It will run no model itself: it will talk to a server such as neurobrix serve. |
Containers (.nbx) |
One container per model, in one open format: its graph, its weights and everything needed to run it. | A container holds no code. |
| Hub (neurobrix.org) | The catalogue: it publishes the containers, the releases and this documentation, and serves the downloads. | It runs no model. |
Tensor code lives only in nbxtensor. The engine decides what runs, where and in which order; nbxtensor does the tensor work. A new kernel, dtype or memory mechanism is added to nbxtensor, never to the engine. The engine is a user of nbxtensor, like any other program: it imports the library's functions and does no tensor work of its own.
pip install neurobrix installs the engine together with the exact version of nbxtensor it was released with, and of NBXGrammar where its wheels exist, so the bricks you get always match.© 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