Machine Heart

Nov 6, 2026 · Artificial Intelligence

S²-MAS: 12M Household Simulation at 37K Tokens/s, 421M YN-mini Beats Jev Model

HKU researchers introduce S²-MAS, a framework simulating 12 million households using 14 GPUs in 4 hours, featuring a 500M persona library, lightweight YN-mini decision model (421M params) achieving 37K tokens/s on a single GPU, outperforming larger models like Jev and GLM in consumer finance choice accuracy, and enabling granular analysis of heterogeneous household responses to policy changes.