**[DeepSeek Collaborates with Tsinghua and Peking University on Groundbreaking Paper: Advancing Intelligent Agent Infrastructure to Break Through Agent Reasoning I/O Bottlenecks]** Reported by Jinse Finance, a groundbreaking paper was released ahead of the launch of DeepSeekV4. DeepSeek, in collaboration with research teams from Peking University and Tsinghua University, proposed a novel large model inference system called DualPath. This system introduces a "Dual-Path KV-Cache Loading" mechanism, cleverly utilizing idle network bandwidth within clusters to boost offline inference throughput for Agentic large models by up to 1.87 times and average online service throughput by 1.96 times. Currently, this research has undergone large-scale validation on clusters containing up to 1,152 GPUs, supporting top-tier large models such as DeepSeek-V3.2660B. (AI Cambrian)