SemiAnalysis Posts on vLLM Kimi Throughput Gains

The official SemiAnalysis account posted replies crediting vLLM optimizations with raising high-concurrency throughput by more than 6x in 2 weeks for the Kimi K3 2.8T model. The posts reference a chart of token throughput per chip versus P90 interactivity and state that the firm's AgentX benchmark uncovered issues in multi-turn long-context tasks. The linked article is titled AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing? It lists a $3 Million USD dataset open sourced along with 1 Mil+ Context Length, Multiturn, Sub Agents, 95%+ KVCache HitRate, GB300 NVL72, MI355, and B200.