Preprint reports LUT-based method for running 70B LLMs on one A100

A new arXiv preprint describes FluxBin, an algorithm-and-kernel design for ultra-low-bit LLM inference. The authors report that its post-training quantization and optimized CUDA kernel reduce floating-point work through lookup tables and scale fusion. Across evaluated architectures, they report up to 5.92 times speedup, up to 10.19 times energy savings, and comparable accuracy to heavily fine-tuned methods. They also report a fourfold memory reduction that enables deployment of 70B-scale models on a single A100 GPU. The work is a preprint, and the reported results have not been independently verified.