Frontier models shouldn’t need frontier hardware.
NoxQuantum builds computationally efficient AI: smaller models, faster inference and, over time, better LLM architectures. Built from Africa.
Our researchOur thesis
Large AI models should run on much less hardware. We work on model compression and faster inference, with the longer-term goal of building better LLM architectures.
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
13.48 GB: original Llama-2-7B package
2.46 GB: our NEC result. 5.485× compression and 81.77% less storage.
Compute is a constraint we know first-hand. Our work focuses on smaller model representations, faster inference and, over time, more efficient LLM architectures. We’re building from Africa for the hardware people actually have.
Our compression result
Compact models
NEC compresses all 32 layers of Llama-2-7B using tensor networks and sparse corrections. The complete model package is 5.485× smaller, with 50.7% fewer stored coefficients.
Construction uses calibration data and analytical solves, with zero gradient-based healing. Relative accuracy retention across evaluated tasks is 71.6–88.9%.
Read the results- Original model coefficients
- 6,738,417,664
- NEC stored coefficients
- 3,323,032,892