Community
NVIDIA's 30B speech LLM pruned from 67.5GB to 10GB without retraining
A community post reports the pruned model still transcribes at 5.16% WER, against 4.12% for the unpruned teacher. The claim comes from a Reddit post and has not been independently verified.