ZiPu Collaborates with Tsinghua to Launch New Networking Architecture, Increasing GLM Inference Throughput by 15%

Chinese artificial intelligence company ZhiPu disclosed on Wednesday that its joint venture with Yù Xùn Network and Tsinghua University has completed the large-scale implementation of the next-generation network architecture ZCube. Without adding any additional GPU hardware, it has increased the inference throughput of the flagship model GLM-5.1 by 15% and significantly reduced infrastructure costs.

The ZCube architecture has been deployed in the online production cluster of GLM-5.1, with the testing scenario focusing on computation-intensive code generation tasks. Official data shows that with the GPU computing power, software stack, and upper-layer applications remaining unchanged, ZCube has reduced the capital expenditure for switches and optical modules by 33%, while reducing the P99 metric for the first token delay (TTFT) by 40.6% — meaning, under high concurrent stress, the waiting time for 99% of user requests to receive the first byte response is shortened by more than 40%.

ZhiPu stated that the 15% throughput increase directly corresponds to a proportional increase in the number of API platform requests that can be processed per second, providing a more stable user experience during peak traffic periods without additional hardware purchase expenditures.

Eliminating the Spine Layer, Full Interconnection Design

According to technical materials, the core innovation of ZCube lies in the reconstruction of the traditional three-tier topology of data center networks. This architecture fundamentally eliminates the common network congestion problems in ultra-large scale inference clusters by canceling the Spine (spine) switching layer, grouping Leaf (leaf) switches, and achieving full interconnection between groups.

In the traditional "access layer—aggregation layer—core layer" architecture, the Spine layer is the concentrated source of bandwidth bottlenecks and single points of failure. ZCube bypasses this bottleneck by directly fully interconnecting Leaf switches, making the communication paths between GPUs shorter and the bandwidth utilization rate higher, thereby releasing the inference potential locked by network latency without increasing computing resources.

This design was validated before ZCube entered large-scale production deployment, having been tested in simulation environments and real test platforms, with early academic results published at the CCF China Networking Conference.

GLM-5.1: Code Capabilities Amongst the World's Best

The GLM-5.1 served by ZCube, officially open-sourced by ZhiPu in April 2026, is the company's current flagship model with the highest technical level. GLM-5.1 uses a mixture of experts (MoE) architecture, with a total parameter count of 744B, activating approximately 40 to 44B parameters per inference, and supports a context window of 200K tokens.

In the most referential code evaluation benchmark globally, SWE-Bench Pro, GLM-5.1 set a new record with a score of 58.4, surpassing OpenAI GPT-5.4 (57.7 points) and Anthropic Claude Opus 4.6 (57.3 points), becoming the first Chinese open-source model to surpass top proprietary products on this benchmark.

In terms of long-duration task capabilities, GLM-5.1 has the ability to operate autonomously for more than 8 hours in a single task, completing the entire cycle of code writing, debugging, and iteration without supervision. ZhiPu's public case studies show that the model has increased vector database query throughput from approximately 3,500 QPS to 21,500 QPS through 655 rounds of autonomous iteration.

The GLM series has been officially accessed by at least 18 companies, including ByteDance TRAE, Alibaba Qoder, Tencent CodeBuddy, Baidu Intelligent Cloud, Meituan, and Kuaishou. ZhiPu disclosed in its first financial report after listing that 9 out of the top 10 Chinese internet companies have deeply integrated GLM.

It is worth noting that with the release of GLM-5.1, ZhiPu simultaneously increased the GLM series API prices by 10%, aligning with the global AI pricing trend — in the context of continuously expanding computing investments, major suppliers of large models are shifting from gaining market share with low prices to strengthening profit margins.

The introduction of the ZCube architecture has dual strategic significance in this context: on the one hand, it reduces the unit cost of infrastructure, creating some room for price increases; on the other hand, it establishes engineering barriers in two core service indicators, throughput and response latency, helping to maintain a quality service advantage in the competition with domestic and international rivals.