Sherry

Native low-bit LLM architectures outperform post-training compression

The article argues that post-training compression techniques for large language models are fundamentally flawed. It explains that these methods, which attempt to reduce model size by quantizing weights after training, f…