New 4/6 quantization method boosts LLM accuracy with adaptive scaling

Researchers have developed a new quantization method called Four Over Six (4/6) to improve the accuracy of low-precision numerical formats like NVFP4 for large language models. This technique adaptively scales blocks to smaller FP4 values, reducing quantization error, particularly for near-maximal values. Experiments with the Nemotron 3 Nano 30B-A3B model architecture showed that 4/6 brings training loss closer to BF16 compared to existing NVFP4 methods, with minimal computational overhead. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Improves efficiency of LLMs by reducing memory usage and increasing speed with minimal accuracy loss.

RANK_REASON Academic paper detailing a new method for model quantization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• NVFP4
• Four Over Six
• Nemotron 3 Nano 30B-A3B
• BF16
• arXiv

• paper
• infra