Zhipu AI Launches Flagship Model Powered Entirely by Domestic Chips

Zhipu AI, a Chinese artificial intelligence organisation, has launched its GLM-5.3-Flash frontier model, according to reports from Pandaily. This model operates entirely on domestically produced AI chips. It has gained attention across both benchmark and usage rankings, indicating a notable development in China's domestic compute capabilities. One hundred thousand locally made accelerators reportedly support all of its online inference traffic.

Credit: Cloudinary

The GLM-5.3-Flash model secured the tenth position on the AAII leaderboard, published by benchmarking firm Artificial Analysis. This placement positioned it ahead of DeepSeek's V4 Pro Max model. Zhipu AI stated that the complete workload for the GLM-5.3-Flash model's online requests is serviced by the array of 100,000 domestically manufactured chips. The performance data suggests the efficacy of the homegrown hardware stack.

This model, initially designated "Ox Alpha," saw its release earlier in Aug. Within a week of its introduction, it achieved the highest usage rankings on OpenRouter, an AI model routing platform. Zhipu AI positions GLM-5.3-Flash as a high-performance model, designed to be efficient enough for operation at scale within China's domestic compute infrastructure. And this rapid uptake on OpenRouter suggests its practical appeal.

Zhipu AI has not publicly disclosed the supplier of the chips utilised in its new model. CNBC has reported that analysts conducting observations suspect the hardware to be from Huawei's Ascend series. Ivan Lin, an analyst at research firm Counterpoint, has noted a trend among Chinese AI model developers. They are increasingly directing investment towards AI servers and compute infrastructure that relies upon domestically produced chips.

The introduction of GLM-5.3-Flash represents a clear indication from at least one of China's significant model laboratories. They appear to believe the domestic supply chain is now capable of managing flagship workloads. Zhipu AI has been identified as one of China's "AI tigers," showing aggression in expanding its compute resources. So, running a frontier-class model on homegrown accelerators daily communicates both performance capabilities and the availability of these specific chips.

This development also aligns with a wider industry movement. Foreign frontier chips have faced export controls, prompting Chinese model builders to accelerate efforts. These developers have focused on optimising their software to function effectively with the hardware they are able to acquire locally. A model demonstrating top benchmark results while operating on domestic chips offers evidence. It suggests that software efficiency can help mitigate existing gaps in hardware capabilities.

For Zhipu AI, the "Flash" product line embodies a strategic approach. This strategy involves combining model quality with practical scalability. With its extensive deployment of 100,000 chips, the GLM-5.3-Flash model aims to demonstrate a China-based laboratory's ability. It seeks to deliver competitive frontier performance without relying on imported accelerator technology. But the long-term implications for the global AI chip market remain to be observed.

• Zhipu AI's GLM-5.3-Flash model operates exclusively on 100,000 domestically produced AI chips.

• The model ranked tenth on the Artificial Analysis AAII leaderboard and secured the top position on OpenRouter usage rankings shortly after its release as "Ox Alpha."

• Analysts, as reported by CNBC, suspect Huawei's Ascend series chips are the unconfirmed hardware provider for the model.

• This development signals a growing confidence in China's domestic supply chain for carrying significant AI workloads, amidst international export controls on foreign chips.

• The strategy underscores an effort to combine model quality with practical scalability using homegrown technology.

Source: Pandaily