DualQuant Quantizes LLMs to 4-bit Without Calibration Data

Researchers at d-Matrix unveiled DualQuant, a post-training quantization framework that compresses LLM weights and activations to 4-bit without a calibration dataset. On Llama-3.1-8B running on an RTX 4090 it delivered 2.61x memory reduction and up to 2.46x throughput speedup over BF16 while matching SmoothQuant accuracy.

• DualQuant quantizes weights and activations to 4-bit with no calibration data
• Llama-3.1-8B on RTX 4090: 2.61x less memory, up to 2.46x faster than BF16
• Targets MXFP4 and NVFP4 formats used by AMD, Intel and Nvidia accelerators
• Paper posted to arXiv on October 9, 2026