DeepSeek expands national compute access as Chinese model activity continues
China’s AI compute landscape saw new DeepSeek deployment infrastructure, continued model releases and downloads, and rising attention to the cost and availability of computing power.
• The National Supercomputing Internet launched DeepSeek V4 Pro and DeepSeek Harness, offering one-stop model deployment and development through a domestically developed resource pool combining supercomputing and intelligent computing capacity at the 100,000-card level.
• Alibaba’s Qwen open-source model family was reported to have surpassed 3 billion cumulative global downloads over six months, including 2.045 billion downloads from the Hugging Face Hub.
• Zhipu released GLM-5.3, which Honor integrated into its YOYO Claw assistant; Honor said the model’s programming ability increased 50% from the previous generation and that it ranked first among open-source models on several public benchmarks.
• Tencent’s second-quarter capital expenditure reached 52.8 billion yuan, while an analysis said computing-power resale generated more than 30% profit and identified HBM and advanced packaging as supply-chain bottlenecks.
• Apple was reportedly developing a China-specific AI model with Alibaba’s partnership and support, according to an item in IT Home’s morning brief.
• DeepSeek Harness drew rapid developer interest but was described in related reports as demanding close user oversight and having a high barrier for ordinary users.