Gemma 3 4B — NMX
Exploring neural models beyond the weights.
NMX started with a simple question:
What if we looked at neural networks not just as models to run, but as systems we can understand, rethink, and optimize?
This repository is part of NMX (Neural Model eXchange), an independent AI research project focused on neural model representation, inference optimization, and the relationship between AI architectures and the hardware they run on.
This release brings Gemma 3 4B into the NMX project.
Why NMX?
The AI ecosystem is moving fast. Models are getting more capable, inference engines are getting smarter, and hardware is evolving in parallel.
But there is still a lot to explore in the space between the model itself and the machine executing it.
How should neural models be represented?
Where does inference efficiency actually come from?
How much can we gain from better execution strategies rather than simply scaling up model size?
These are the kinds of questions that drive NMX.
About this model
Base model: Gemma 3 4B
Project: NMX — Neural Model eXchange
Focus: Neural networks, model engineering, and inference research
This repository explores the use of Gemma 3 4B within the NMX project. The exact modifications, training procedure, and inference optimizations depend on the artifacts and implementation provided in this release.
NMX is an evolving project. The goal is to document what works, investigate what does not, and build a foundation for reproducible experiments.
The bigger picture
NMX is not about claiming that every existing AI system needs to be replaced.
It is about asking better engineering questions.
A model is more than a parameter count. Real-world performance also depends on memory movement, numerical precision, runtime overhead, hardware utilization, and the way the entire inference pipeline is designed.
Sometimes, the interesting breakthrough is not a bigger model.
It is a smarter way to make the model work.
Research status
NMX is an ongoing independent research project. Performance claims and architectural improvements should be supported by reproducible benchmarks before being considered validated.
As the project develops, this page will document its implementation, experiments, limitations, and results.
Built to explore. Measured to improve. Open to challenge.
Disclaimer: Gemma is a model family developed by Google. NMX is an independent project and is not affiliated with or endorsed by Google.