The S-X Quantization Family: S-X8 v4.3 Format, .sx8v43 Container Specification and Methodology

This paper presents S-X8 v4.3, a weight quantization format for large language models that stores 7.50 bits per weight — over 11% fewer bits than the widely used Q8_0 (8.50 bpp) — while delivering FP16-level quality (perplexity within +0.2% of the unquantized model on wikitext-2, inside statistical noise) and 9× less perplexity loss than Q8_0 (+0.26% vs. +2.40% over FP16). On Qwen3.5-4B, the format reduces the text payload by 11.6% and the complete model by 15%, and runs decoding faster than Q8_0 in real-world use (63.8 vs. 40–54 tokens/s on an RTX 5060 Ti). The format is byte-aligned, uses only ~9–10 ALU operations per weight for decoding (no shared memory, no shuffles, no tensor-core dependency), making it portable across GPU architectures and backends. Quality is validated with a strict protocol (perplexity, Winogrande_s, HellaSwag, ARC-Challenge, MMLU) against FP16 and Q8_0 on the same machine, and the format is integrated as a native type in a llama.cpp fork, including a tensor-core matrix-multiply kernel for prompt processing. This record includes the container specification (.sx8v43) and the methodology of the S-X quantization family, and covers any derivative of this methodology at any bit-width or platform.