TileQuant quantization compression

Tag "TileQuant quantization compression" — all articles

062026-11Two arXiv papers propose new methods for compressing large modelsarXiv published two studies on large-model compression on Nov 6, 2026. TileQuant uses hardware-aligned 2D weight tiles as mixed-precision units and achieves 2.66x generation throughput without training; the other work exploits the degrees of freedom in equivalence classes of network parameterizations to cut Zstandard-compressed checkpoint size by up to 23.4% on SmolLM2 and Qwen3. Both target deployment costs and remain preprints.arXiv cs.LG