Quantum-floor compression: Achieving GPT-4 capability at 1/120th the model size [pdf]

Read more here: External Link