DeepSeek collaborates with Tsinghua University to publish a heavyweight hardcore paper: Fueling the infrastructure of intelligent agents at the bottom level, breaking through the I/O bottleneck of agent reasoning.

On the eve of DeepSeekV4's release, a hardcore heavyweight paper was released online. The DeepSeek team, in collaboration with research teams from Peking University and Tsinghua University, proposed a new large model inference system called DualPath. This system cleverly utilizes the idle network bandwidth in the cluster by introducing the "DualPath KV-Cache Loading" mechanism, increasing the offline inference throughput of Agentic large models by up to 1.87 times, and the online service throughput by an average of 1.96 times. Currently, this research has been successfully validated on a cluster with up to 1152 GPUs, supporting top large models such as DeepSeek-V3.2660B.