Frontier models shouldn’t need frontier hardware.
NoxQuantum is a research company built from Africa, developing quantum and quantum-inspired algorithms for hard computational problems, starting with compact AI models.
Our approachOur thesis
Large AI models should run on much less hardware without losing the capabilities that matter. We want to find out whether ideas from quantum information can help.
Why Africa
The African Union reports that most African countries lack powerful GPUs in their universities and research institutions.
Source: African Union, Continental Artificial Intelligence Strategy (July 2024), p. 48
14 GB: a 7B-parameter model with 16-bit weights
1.4 GB: what a tenth of that would be. That’s our target, not something we’ve achieved.
Most frontier models can’t be run there because the hardware isn’t available. A model that needs a tenth of the memory could be used in classrooms, clinics and field systems. That’s why Nox is starting in Africa, and why we plan to stay here as we grow. Cheaper models would matter elsewhere too.
The first problem
Compact models
A large weight matrix can be stored as a chain of small tensors. We want to find out whether this, and related ideas from quantum information, can beat pruning, quantization and distillation.
These numbers only count what has to be stored. We don’t yet know whether a trained model’s weights can be squeezed that far without losing what matters. That’s what we’re testing.
How we work- One 4096 × 4096 matrix
- 16,777,216 numbers
- Four cores, bond dimension 16
- 34,816 numbers