ResearchBenchmarksMatrix Multiplication Latency on Llama-3 70BFollow293.82Kernel Latency (µs)CodeGEMM214.1528751.90641,289.661,827.4136Dec 19, 2025Evaluation ResultsMethodMethodLinksKernel Latency (µs)CodeGEMMconfig=m1v4g128config=m1v4g1282025.12293.82LUTGEMMquantization=q2-g128quantization=q2-g1282025.12299.9CodeGEMMconfig=m2v8g128config=m2v8g1282025.12373.49QuIP#config=e8pconfig=e8p2025.12403.59QTIPconfig=r2config=r22025.12477.04AQLMconfig=2x8config=2x82025.12674.67cuBLASprecision=fp16precision=fp162025.121,111.36AQLMconfig=1x16config=1x162025.122,285.5