๊ฐ€์ž… ํ›„ ์ดˆ๋Œ€ ๋งํฌ๋ฅผ ๊ณต์œ ํ•˜๋ฉด ๋™์˜์ƒ ์žฌ์ƒ ๋ฐ ์ดˆ๋Œ€ ๋ณด์ƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

LightSeek Foundation
@lightseekorg
LightSeek is the creator of TokenSpeed, TorchSpec, and SMG โ€” building next-generation AI infra systems
๊ฐ€์ž… October 2025
2 ํŒ”๋กœ์ž‰ ์ค‘    3.3K ํŒฌ
331 output tokens/s per GPU on @Kimi_Moonshot K3 NVFP4, powered by TokenSpeed โšก๏ธ Running the full @SemiAnalysis_ AgentX trace on 32 @NVIDIAAI GB300 GPUs, TokenSpeed achieves 1.92ร— the output throughput per GPU of the standard eight-GPU DCP8 baseline. Cache hit rate and workload balance are the foundations of high-throughput agentic inference. Our latest blog explains how independent DWDP prefill, efficient DEP decode, and shared caching work together to keep long-running agent sessions moving. ๐Ÿ‘‡ ๐Ÿ“– โญ๏ธ
๋” ๋ณด๊ธฐ