AI 导读
Jalapeño 每 kW 吞吐量提升 1.9 倍，端到端延迟降低 1.7 倍，token 延迟降低 2.7 倍。 在此前最佳解码速度下，其吞吐优势跃升至 53.7 倍：22,935 对 427 mixed tokens/s/kW。https://t.co/01l91X6W4J
正文
Jalapeño delivers 1.9x more throughput per kW while cutting end-to-end latency 1.7x and token delay 2.7x.
At the previous best decoding speed, its throughput advantage jumps to 53.7x: 22,935 vs. 427 mixed tokens/s/kW. https://t.co/01l91X6W4J
来源：@rohanpaul_ai · x.com