Inference-time speed innovations allow you to take distilled or quantized models and make them 100x faster, which paradoxically makes them far more intelligent than the original exported model from the data center, creating a massive leak of capability.