model-moe

Ling-3.0’s inference bottleneck, a 124B Mixture-of-Experts model with just 5.1B active parameters per token, redefines AI supply chain economics & latency.