Listwise passage reranking often achieves strong ranking performance, but it incurs substantial computational cost. We propose FLASH, an efficient listwise reranking framework based on semantic token compression and single-step listwise scoring. FLASH represents each query–passage pair with a small set of query-aware semantic tokens and jointly scores all candidates in a single decoder step, enabling query-conditioned joint scoring without expensive token-level listwise processing. We further introduce multi-cardinality training, which allows a single model to support different semantic-token budgets at inference time. We also propose FLASH-LITE, a lightweight variant with offline passage encoding for low-cost online inference. Across TREC and BEIR benchmarks, FLASH improves average ranking effectiveness at comparable online latency and slightly lower FLOPs, while FLASH-Lite substantially reduces query-time cost with a modest effectiveness trade-off. FLASH-LITE enables offline passage encoding, substantially reducing online query-time cost while maintaining competitive reranking performance. Code is provided in the supplementary material.