TileQuant preprint reports 2.66x generation throughput using mixed precision

Researchers introducing TileQuant report generation throughput of 2.66x the original model's in experiments with a training-free method that assigns different numerical precision to small blocks of language model weights. Calibration identifies the most sensitive blocks, which remain at 16-bit precision while the rest use 8-bit. The execution kernel combines integer and 16-bit computation to reduce memory traffic. The preprint reports preserved quality across language modeling, reasoning, long-context retrieval and question answering, with results surpassing quantized and original models on multiple tasks.