China's Ministry of Industry and Information Technology (MIIT) has released a new directive to deepen the application of artificial intelligence across industries, particularly targeting small and medium-sized enterprises. This policy move is expected to fuel further growth in AI training, reasoning, and agent-based applications.
On July 27, the MIIT issued the "Guidance on Cultivating 'Small, Fast, Light, and Precise' Digital Products and Services (2026 Edition)". The document encourages the deep integration of AI technology into products, pushing for a shift from rule-based or model-driven logic reasoning and decision support towards multi-modal perception, predictive analytics, and autonomous execution. The goal is to accelerate the deep integration of AI within small and medium-sized enterprises.
This follows the State Council's earlier "Opinions on Deepening the 'AI+' Action", which set a clear target for 2027: to achieve widespread, deep integration of AI across six key sectors, with a penetration rate of over 70% for next-generation intelligent terminals and agents. The core AI economy is expected to scale rapidly, AI's role in public governance will be significantly enhanced, and the open cooperation system for AI will be continuously improved.
In 2025, China's core AI industry has already surpassed a scale of 1.2 trillion yuan, with the application rate of AI technology in enterprises above a designated size exceeding 30%. The country's vast market, complete industrial system, and rich application scenarios provide strong support for AI technology innovation, product iteration, and large-scale deployment. As the "AI+" initiative deepens, technological potential, market advantages, and institutional strengths are expected to further translate into developmental success.
According to the latest data from OpenRouter, the total global call volume for large AI models last week was 58 trillion tokens, a sequential decline of 6.75%. Chinese large models accounted for 33 trillion tokens of this total, down 8.61% from the previous week. In comparison, the weekly call volume for US large models was 2.34 trillion tokens, a drop of 67.86%. This means China's large model call volume is now 14.1 times that of the US and has held the top global position for thirteen consecutive weeks.
Last week, all five of the top global call volume leaders were Chinese AI large models. XIAOMI-W (HK:01810) MiMo-V2.5 rose to the number one spot with a weekly call volume of 10.5 trillion tokens, a 12% increase from the previous week. Previously, TENCENT (HK:00700) Hy3 held the top position for two consecutive weeks during its free trial period. After the free version was removed from OpenRouter last week, its weekly call volume fell 85% to 1.76 trillion tokens, dropping it to ninth place. However, the free trial successfully completed developer outreach, and the call volume for the commercial version of Hy3 surged to 3.94 trillion tokens, ranking third with a sequential increase of over 999%.
XIAOMI-W's MiMo-V2.5 series began public testing on April 23, including models MiMo-V2.5, V2.5-Pro, V2.5-TTS Series, and V2.5-ASR. It features enhanced reasoning, a more stable agent, longer context windows, better instruction following and fuzzy instruction understanding, and improved full-modal perception and comprehension.
The second-ranked model was DeepSeek-V4-Flash, with a weekly call volume of 6.37 trillion tokens. This version is designed for daily text processing, high-throughput inference, and cost-sensitive deployment scenarios. DeepSeek offers a high cost-performance ratio, with the Flash version priced at $0.14 (input) / $0.28 (output), compared to $1.74 (input) / $3.48 (output) for the Pro version.
TENCENT Hy3 was officially released on July 6. According to the company, Hy3 is a model that integrates fast and slow thinking, using a Mixture of Experts (MoE) architecture with 295 billion total parameters and 21 billion activated parameters, supporting a 256K context length. Compared to the preview version, its intelligence level is significantly stronger than other models of a similar size, comparable to flagship models two to five times its size, with substantial improvements in overall stability and cost-effectiveness.
CITIC Securities holds a long-term bullish view on domestic computing power, stating that "computing power is national power," and the market should focus on the demand and breakout of local computing power. Currently, domestic internet companies' capital expenditures in the AI field are significantly lower than their overseas counterparts. With rising willingness to invest and the massive advantages of Chinese companies in applications, domestic demand for computing power is expected to grow faster than overseas markets by 2030.
Dongguan Securities pointed out that on the demand side, the ongoing volume expansion of AI training, reasoning, and agent applications, along with upward revisions in overseas cloud CAPEX, will maintain high prosperity in computing infrastructure construction. This will continue to drive demand for storage, CPUs, power semiconductors, and semiconductor materials, with some sub-categories already showing supply-demand tightness and price increases. On the supply side, key areas such as semiconductor equipment, high-end GPUs, storage chips, and advanced packaging are still in an accelerated phase of domestic substitution, with a strong push for an independent and controllable supply chain.
Among the key related stocks, BIDU-SW (HK:09888) has seen its Wenxin 5.0 official version surpass models like Gemini-2.5-Pro and GPT-5-High in a comprehensive evaluation of over 40 authoritative benchmarks, securing a top-tier international position. Its image and video generation capabilities are on par with specialized vertical models, and overall it is a global leader. The monthly active users of the Baidu Wenxin Assistant have already exceeded 200 million.
BABA-W (HK:09988) released its flagship large model, Qwen3.8 Max Preview, on July 19. With a massive 2.4 trillion parameters, the company describes it as "one of the most powerful models today, comparable to leading frontier AI models," ranking second globally only behind Anthropic's Fable 5.
MINIMAX-W (HK:00100) officially open-sourced its MiniMax M3 model on June 15. The model has 428 billion total parameters and 23 billion activated parameters. It is described as achieving "industry-leading levels" in coding and agent evaluations, with the ability to autonomously decompose tasks, call tools, and perform multi-step reasoning. The code it writes is reportedly designed to be directly deliverable, not just "runnable but requiring human modification."
Finally, Z.AI (HK:02513) has released its new flagship large model, GLM-5.2. Built for long-context tasks, it supports a usable million-token context length. In international comprehensive evaluations, GLM-5.2 ranks among the top global open-source models, achieving internationally advanced levels in core capabilities like programming and long-context tasks. This marks a new breakthrough in the key technical capabilities of China's domestic large models.
Comments