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%)

🌐 链接: https://www.reddit.com/r/LocalLLaMA/comments/1xibujk/pruned_nvidias_new_30b_speech_llm_from_675gb_to/
🏷️ 分组: Reddit｜AI·自托管·VPS
🕒 时间: 2026-10-29 13:54:38
Reddit
From the LocalLLaMA community on Reddit: pruned NVIDIA's new 30B speech LLM from 67.5GB to 10GB with zero retraining, it still…
From the LocalLLaMA community on Reddit: pruned NVIDIA's new 30B speech LLM from 67.5GB to 10GB with zero retraining, it still…
Explore this post and more from the LocalLLaMA community