pruned NVIDIA's new 30B speech LLM from 67.5GB to 10GB with zero retraining, it still transcribes at 5.16% WER (teacher: 4.12%) · AI Intel