Explainer: How 30B-parameter models activate only 3B per token

The post shared a technical explainer on how a 30B-parameter model can activate just 3B parameters per token while using the full model's capacity. It covers dense vs. MoE models and implications for throughput, memory, and serving complexity.