The film is three and a half minutes long. On YouTube you can pick the audio track: English, français, español or العربية.

Why we are building NeuroBrix

For most people, artificial intelligence has one face: a chat assistant, rented by the minute, running in someone else's data center. That face is real and useful, but it is a small part of what neural networks do. They segment cells under a microscope, transcribe a consultation, find a defective part on a production line, upscale an image before printing, predict the shape of a protein.

None of those jobs needs one giant model that does everything. Each needs the right model, running where the data lives, on the hardware the organisation already owns. That is the gap NeuroBrix is built for.

An ecosystem, not a product

NeuroBrix is made of bricks. You can take the whole stack, or only the brick you need.

  • The engine. One runtime for text, image, audio and video models, loaded from a single container format.
  • nbxtensor. Our own tensor library. It is what the engine computes with, and it is published as a library of its own, so you can build your own tools on it.
  • Kernels. Written for each family of processors, so the hardware you have is used for what it can actually do.
  • NBXGrammar. A constrained-decoding engine in Rust: the model answers in exactly the shape you asked for (JSON, a tool call, a grammar of your own).
  • Certified containers. Every model ships in a container that is checked to behave the same way on every machine.
  • The hub. The library where those containers wait, ready to run: neurobrix.org.
  • NBX Code (coming soon). A coding agent that runs on your machines and sends nothing outside.

Three examples

  • A laboratory segments its cells with Cellpose and predicts protein structures with ESMFold, and its data never leaves the building.
  • A hospital transcribes consultations with Whisper and outlines an organ on a scan with MedSAM.
  • A print shop enlarges an image with Real-ESRGAN before a large-format run, on the card it already has.

The hardware you already own

The industry moves fast, and every new generation of processors is good news. But a processor does not stop computing the day its successor arrives. One measured example: on NVIDIA V100 Tensor Cores, our own attention kernels run up to 17 times faster, and a 30-billion-parameter model served on two V100s went from 6.7 seconds per token to about a third of a second, with no hardware change.

From a laptop graphics chip to a rack of servers, a new processor is one more brick.

Try it, break it, build with it 

pip install neurobrix
  • Developers: the code is on GitHub, Apache 2.0. A star helps more than you think: github.com/NeuroBrix/neurobrix
  • Testers: run a model on your own machine and open an issue when something breaks. Every report makes the next release better.
  • Companies, partners, investors: if you want to run AI in-house, or build what comes next with us: neurobrix.org/enterprise

Documentation: docs.neurobrix.org